In deep regression, capturing the relationship among continuous labels in feature space is a fundamental challenge that has attracted increasing interest. Addressing this issue can prevent models from converging to suboptimal solutions across various regression tasks, leading to improved performance, especially for imbalanced regression and under limited sample sizes. However, existing approaches often rely on order-aware representation learning or distance-based weighting. In this paper, we hypothesize a linear negative correlation between label distances and representation similarities in regression tasks. To implement this, we propose an angle-compensated contrastive regularizer for deep regression, which adjusts the cosine distance between anchor and negative samples within the contrastive learning framework. Our method offers a plug-and-play compatible solution that extends most existing contrastive learning methods for regression tasks. Extensive experiments and theoretical analysis demonstrate that our proposed angle-compensated contrastive regularizer not only achieves competitive regression performance but also excels in data efficiency and effectiveness on imbalanced datasets.

Enabling object detectors to recognize out-of-distribution (OOD) objects is vital for building reliable systems. A primary obstacle stems from the fact that models frequently do not receive supervisory signals from unfamiliar data, leading to overly confident predictions regarding OOD objects. Despite previous progress that estimates OOD uncertainty based on the detection model and in-distribution (ID) samples, we explore using pre-trained vision-language representations for object-level OOD detection. We first discuss the limitations of applying image-level CLIP-based OOD detection methods to object-level scenarios. Building upon these insights, we propose RUNA, a novel framework that leverages a dual encoder architecture to capture rich contextual information and employs a regional uncertainty alignment mechanism to distinguish ID from OOD objects effectively. We introduce a few-shot fine-tuning approach that aligns region-level semantic representations to further improve the model's capability to discriminate between similar objects. Our experiments show that RUNA substantially surpasses state-of-the-art methods in object-level OOD detection, particularly in challenging scenarios with diverse and complex object instances.

Zhigen Li, Jianxiang Peng, Yanmeng Wang, Yong Cao, Tianhao Shen, Minghui Zhang, Linxi Su, Shang Wu, Yihang Wu, YuQian Wang, Ye Wang, Wei Hu, Jianfeng Li, Shaojun Wang, Jing Xiao, Deyi Xiong. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Co-speech gestures are essential to non-verbal communication, enhancing both the naturalness and effectiveness of human interaction. Although recent methods have made progress in generating co-speech gesture videos, many rely on strong visual controls, such as pose images or TPS key-point movements, which often lead to artifacts like blurry hands and distorted fingers. In response to these challenges, we present the Implicit Motion-Audio Entanglement (IMAE) method for co-speech gesture video generation. IMAE strengthens audio control by entangling implicit motion parameters, including pose and expression, with audio inputs. Our method utilizes a two-branch framework that combines an audio-to-motion generation branch with a video diffusion branch, enabling realistic gesture generation without requiring additional inputs during inference. To improve training efficiency, we propose a two-stage slow-fast training strategy that balances memory constraints while facilitating the learning of meaningful gestures from long frame sequences. Extensive experimental results demonstrate that our method achieves state-of-the-art performance across multiple metrics. Project Page.

Recent studies have demonstrated that many layers are functionally redundant in large language models (LLMs), enabling model compression by removing these layers to reduce inference cost.While such approaches can improve efficiency, indiscriminate layer pruning often results in significant performance degradation.In this paper, we propose GRASP (Gradient-based Retention of Adaptive Singular Parameters), a novel compression framework that mitigates this issue by preserving sensitivity-aware singular values.Unlike direct layer pruning, GRASP leverages gradient-based attribution on a small calibration dataset to adaptively identify and retain critical singular components.By replacing redundant layers with only a minimal set of parameters, GRASP achieves efficient compression while maintaining strong performance with minimal overhead.Experiments across multiple LLMs show that GRASP consistently outperforms existing compression methods, achieving 90% of the original model's performance under 20% compression ratio.The source code is available at https://github.com/LyoAI/GRASP.

Contextual biasing is essential for addressing scenario-specific challenges in End-to-End (E2E) Automatic Speech Recognition (ASR) systems. Prior contextual E2E ASR methods, such as the contextual bias with CPP Network, have utilized bias CTC loss for explicit supervision of bias tasks, However, the alignment between the ASR and bias tasks has been largely neglected. This paper introduces a novel contextual biasing approach that employs the Multi-label Synchronous Output CTC (MCTC) algorithm to enhance the synchronization between ASR and bias task outputs. Furthermore, we propose an enhancement to phrase-level contextual representation. Our proposed method demonstrates significant improvements in Word Error Rate (WER). Specifically, our experiments reveal a 15.0% reduction in WER on the Librispeech-960 dataset compared to the CPPN method, with an impressive 29.0% reduction in WER for context phrases.

Contextual automatic speech recognition (ASR) plays an increasingly important role in addressing the long-tail issues of general ASR. In the past, contextual ASR mainly focused on phrase-level discussions, providing a convenient way to handle biasing phrases. This paper introduces a new contextual network for extracting context tokens, with a focus on optimizing contextual knowledge using token-level information. We enhance the fusion of context information and ASR acoustic features, recognizing that textual knowledge inherently represents a distinct modality compared to acoustic information. More importantly, we propose a creative approach to address the challenge of the increasing size of expanding token list compared to phrase list. Our method is tested on the LibriSpeech and AISHELL-2 datasets, the results demonstrate a 31.0%/23.3% Word Error Rate (WER) reduction on LibriSpeech and a 26.9% reduction Character Error Rate (CER) on the named entity (NE) set from AISHELL-2.

Recently, the field of Text-to-speech synthesis has been predominantly characterized by end-to-end models, with the quality of speech generated by these models becoming increasingly comparable to that of human speech. In this work, we propose a Lightweight and Efficient Text-to-speech model, a fast end-to-end framework based on EfficientTTS 2 with fully differentiable. We utilize Fast Linear Attention with a Single Head instead of the standard stacked Transformer, which decreases computational complexity and reduces parameters. Additionally, we improve a network architecture ConvWaveNet to further decrease model parameters, and accelerate inference through a multi-stream inverse short-time Fourier Transform generator. These improvements significantly reduce model parameters and increase inference speed, thereby achieving the objectives of faster inference and lightweight modeling. Experimental results show that the proposed model achieves speech quality comparable to that of the baseline models, while also offering improved inference speed and reduced model size.

The rapid advancement of large language models (LLMs) has significantly enhanced their reasoning abilities, enabling increasingly complex tasks. However, these capabilities often diminish in smaller, more computationally efficient models like GPT-2. Recent research shows that reasoning distillation can help small models acquire reasoning capabilities, but most existing methods focus primarily on improving teacher-generated reasoning paths. Our observations reveal that small models can generate high-quality reasoning paths during sampling, even without chain-of-thought prompting, though these paths are often latent due to their low probability under standard decoding strategies. To address this, we propose Self-Enhanced Reasoning Training (SERT), which activates and leverages latent reasoning capabilities in small models through self-training on filtered, self-generated reasoning paths under zero-shot conditions. Experiments using OpenAI’s GPT-3.5 as the teacher model and GPT-2 models as the student models demonstrate that SERT enhances the reasoning abilities of small models, improving their performance in reasoning distillation.

Non-autoregressive (NAR) automatic speech recognition (ASR) models predict tokens independently and simultaneously, bringing high inference speed. However, there is still a gap in the accuracy of the NAR models compared to the autoregressive (AR) models. In this paper, we propose a single-step NAR ASR architecture with high accuracy and inference speed, called EffectiveASR. It uses an Index Mapping Vector (IMV) based alignment generator to generate alignments during training, and an alignment predictor to learn the alignments for inference. It can be trained end-to-end (E2E) with cross-entropy loss combined with alignment loss. The proposed EffectiveASR achieves competitive results on the AISHELL-1 and AISHELL-2 Mandarin benchmarks compared to the leading models. Specifically, it achieves character error rates (CER) of 4.26%/4.62% on the AISHELL-1 dev/test dataset, which outperforms the AR Conformer with about 30x inference speedup.

Recently Data-Free Knowledge Distillation (DFKD) has garnered attention and can transfer knowledge from a teacher neural network to a student neural network without requiring any access to training data. Although diffusion models are adept at synthesizing high-fidelity photorealistic images across various domains, existing methods cannot be easiliy implemented to DFKD. To bridge that gap, this paper proposes a novel approach based on diffusion models, DiffDFKD. Specifically, DiffDFKD involves targeted optimizations in two key areas. Firstly, DiffDFKD utilizes valuable information from teacher models to guide the pre-trained diffusion models’ data synthesis, generating datasets that mirror the training data distribution and effectively bridge domain gaps. Secondly, to reduce computational burdens, DiffDFKD introduces Latent CutMix Augmentation, an efficient technique, to enhance the diversity of diffusion model-generated images for DFKD while preserving key attributes for effective knowledge transfer. Extensive experiments validate the efficacy of DiffDFKD, yielding state-of-the-art results exceeding existing DFKD approaches.We release our code at https://github.com/xhqi0109/DiffDFKD.

Generalizability, the capacity of a robust model to perform effectively on unseen data, is crucial for audio deepfake detection due to the rapid evolution of text-to-speech (TTS) and voice conversion (VC) technologies. A promising approach to differentiate between bonafide and spoof samples lies in identifying intrinsic disparities to enhance model generalizability. From an information-theoretic perspective, we hypothesize the information content is one of the intrinsic differences: bonafide sample represents a dense, information-rich sampling of the real world, whereas spoof sample is typically derived from lower-dimensional, less informative representations. To implement this, we introduce frame-level latent information entropy detector(f-InfoED), a framework that extracts distinctive information entropy from latent representations at the frame level to identify audio deepfakes. Furthermore, we present AdaLAM, which extends large pre-trained audio models with trainable adapters for enhanced feature extraction. To facilitate comprehensive evaluation, the audio deepfake forensics 2024 (ADFF 2024) dataset was built by the latest TTS and VC methods. Extensive experiments demonstrate that our proposed approach achieves state-of-the-art performance and exhibits remarkable generalization capabilities. Further analytical studies confirms the efficacy of AdaLAM in extracting discriminative audio features and f-InfoED in leveraging latent entropy information for more generalized deepfake detection.

Diffusion models can synthesize realistic co-speech video from audio for various applications, such as video creation and virtual agents. However, existing diffusion-based methods are slow due to numerous denoising steps and costly attention mechanisms, preventing real-time deployment. In this work, we distill a many-step diffusion video model into a few-step student model. Unfortunately, directly applying recent diffusion distillation methods degrades video quality and falls short of real-time performance. To address these issues, our new video distillation method leverages input human pose conditioning for both attention and loss functions. We first propose using accurate correspondence between input human pose keypoints to guide attention to relevant regions, such as the speaker’s face, hands, and upper body. This input-aware sparse attention reduces redundant computations and strengthens temporal correspondences of body parts, improving inference efficiency and motion coherence. To further enhance visual quality, we introduce an input-aware distillation loss that improves lip synchronization and hand motion realism. By integrating our input-aware sparse attention and distillation loss, our method achieves real-time performance with improved visual quality compared to recent audio-driven and input-driven methods. We also conduct extensive experiments showing the effectiveness of our algorithmic design choices.

Knowledge Graph Completion (KGC) is a task aimed at uncovering the inherent relationships among known knowledge triplets in a Knowledge Graph (KG) and subsequently predicting missing links. Presently, there is a rising interest in inductive knowledge graph completion, where missing links may pertain to previously unobserved entities. Previous inductive KGC methods mainly rely on descriptive information of entities to improve the representation of unseen entities, neglecting to provide effective prior knowledge for relation modeling. To tackle this challenge, we capture prior schema-level interactions related to relations by leveraging entity type information, thereby furnishing effective prior constraints when reasoning with newly introduced entities. Moreover, We employ normal in-batch negatives and introduce schema-guided negatives to bolster the efficiency of normal contrastive representation learning. Experimental results demonstrate that our approach consistently achieves state-of-the-art performance on various established metrics across multiple benchmark datasets for link prediction. Notably, our method achieves a 20.5% relative increase in Hits@1 on the HumanWiki-Ind dataset.

Intent is defined for understanding spoken language in existing works. Both textual features and acoustic features involved in medical speech contain intent, which is important for symptomatic diagnosis. In this paper, we propose a medical speech classification model named DRSC that automatically learns to disentangle intent and content representations from textual-acoustic data for classification. The intent representations of the text domain and the Mel-spectrogram domain are extracted via intent encoders, and then the reconstructed text feature and the Mel-spectrogram feature are obtained through two exchanges. After combining the intent from two domains into a joint representation, the integrated intent representation is fed into a decision layer for classification. Experimental results show that our model obtains an average accuracy rate of 95% in detecting 25 different medical symptoms.

Dance serves as a powerful medium for expressing human emotions, but the lifelike generation of dance is still a considerable challenge. Recently, diffusion models have showcased remarkable generative abilities across various domains. They hold promise for human motion generation due to their adaptable many-to-many nature. Nonetheless, current diffusion-based motion generation models often create entire motion sequences directly and unidirectionally, lacking focus on the motion with local and bidirectional enhancement. When choreographing high-quality dance movements, people need to take into account not only the musical context but also the nearby music-aligned dance motions. To authentically capture human behavior, we propose a Bidirectional Autoregressive Diffusion Model (BADM) for music-to-dance generation, where a bidirectional encoder is built to enforce that the generated dance is harmonious in both the forward and backward directions. To make the generated dance motion smoother, a local information decoder is built for local motion enhancement. The proposed framework is able to generate new motions based on the input conditions and nearby motions, which foresees individual motion slices iteratively and con-solidates all predictions. To further refine the synchronicity between the generated dance and the beat, the beat information is incorporated as an input to generate better music-aligned dance movements. Experimental results demonstrate that the proposed model achieves state-of-the-art performance compared to existing unidirectional approaches on the prominent benchmark for music-to-dance generation. The code and models are available: https://github.com/czzhang179/BADM.

Deploying neural networks on microcontroller units (MCUs) presents substantial challenges due to their constrained computation and memory resources. Previous researches have explored patch-based inference as a strategy to conserve memory without sacrificing model accuracy. However, this technique suffers from severe redundant computation overhead, leading to a substantial increase in execution latency. A feasible solution to address this issue is mixed-precision quantization, but it faces the challenges of accuracy degradation and a time-consuming search time. In this paper, we propose QuantMCU, a novel patch-based inference method that utilizes value-driven mixed-precision quantization to reduce redundant computation. We first utilize value-driven patch classification (VDPC) to maintain the model accuracy. VDPC classifies patches into two classes based on whether they contain outlier values. For patches containing outlier values, we apply 8-bit quantization to the feature maps on the dataflow branches that follow. In addition, for patches without outlier values, we utilize value-driven quantization search (VDQS) on the feature maps of their following dataflow branches to reduce search time. Specifically, VDQS introduces a novel quantization search metric that takes into account both computation and accuracy, and it employs entropy as an accuracy representation to avoid additional training. VDQS also adopts an iterative approach to determine the bitwidth of each feature map to further accelerate the search process. Experimental results on real-world MCU devices show that QuantMCU can reduce computation by 2.2x on average while maintaining comparable model accuracy compared to the state-of-the-art patch-based inference methods.

The audio watermarking technique embeds messages into audio and accurately extracts messages from the watermarked audio.Traditional methods develop algorithms based on expert experience to embed watermarks into the time-domain or transform-domain of signals.With the development of deep neural networks, deep learning-based neural audio watermarking has emerged.Compared to traditional algorithms, neural audio watermarking achieves better robustness by considering various attacks during training.However, current neural watermarking methods suffer from low capacity and unsatisfactory imperceptibility.Additionally, the issue of watermark locating, which is extremely important and even more pronounced in neural audio watermarking, has not been adequately studied.In this paper, we design a dual-embedding watermarking model for efficient locating.We also consider the impact of the attack layer on the invertible neural network in robustness training, improving the model to enhance both its reasonableness and stability.Experiments show that the proposed model, IDEAW, can withstand various attacks with higher capacity and more efficient locating ability compared to existing methods.The code is available at https://github.com/PecholaL/IDEAW.

Voice conversion refers to transferring speaker identity with well-preserved content. Better disentanglement of speech representations leads to better voice conversion. Recent studies have found that phonetic information from input audio has the potential ability to well represent content. Besides, the speaker-style modeling with pre-trained models making the process more complex. To tackle these issues, we introduce an new method named "CTVC" which utilizes disen-tangled speech representations with contrastive learning and time-invariant retrieval.Specifically, a similarity-based compression module is used to facilitate a more intimate connection between the frame-level hidden features and linguistic information at phoneme-level. Additionally, a time-invariant retrieval is proposed for timbre extraction based on multiple segmentation and mutual information. Experimental results demonstrate that "CTVC" outperforms previous studies and improves the sound quality and similarity of converted results.

Existing emotional speech synthesis methods often utilize an utterance-level style embedding extracted from reference audio, neglecting the inherent multi-scale property of speech prosody. We introduce ED-TTS, a multi-scale emotional speech synthesis model that leverages Speech Emotion Diarization (SED) and Speech Emotion Recognition (SER) to model emotions at different levels. Specifically, our proposed approach integrates the utterance-level emotion embedding extracted by SER with fine-grained frame-level emotion embedding obtained from SED. These embeddings are used to condition the reverse process of the denoising diffusion probabilistic model (DDPM). Additionally, we employ cross-domain SED to accurately predict soft labels, addressing the challenge of a scarcity of fine-grained emotion-annotated datasets for supervising emotional TTS training.

This paper introduces INCPrompt, an innovative continual learning solution that effectively addresses catastrophic forgetting. INCPrompt’s key innovation lies in its use of adaptive key-learner and task-aware prompts that capture task-relevant information. This unique combination encapsulates general knowledge across tasks and encodes task-specific knowledge. Our comprehensive evaluation across multiple continual learning benchmarks demonstrates INCPrompt’s superiority over existing algorithms, showing its effectiveness in mitigating catastrophic forgetting while maintaining high performance. These results highlight the significant impact of task-aware incremental prompting on continual learning performance.

Catastrophic forgetting poses a substantial challenge for managing intelligent agents controlled by a large model, causing performance degradation when these agents face new tasks. In our work, we propose a novel solution - the Progressive Prompt Decision Transformer (P2DT). This method enhances a transformer-based model by dynamically appending decision tokens during new task training, thus fostering task-specific policies. Our approach mitigates forgetting in continual and offline reinforcement learning scenarios. Moreover, P2DT leverages trajectories collected via traditional reinforcement learning from all tasks and generates new taskspecific tokens during training, thereby retaining knowledge from previous studies. Preliminary results demonstrate that our model effectively alleviates catastrophic forgetting and scales well with increasing task environments.

Nowadays, singing voice conversion (SVC) has made great strides in both naturalness and similarity for common SVC with a neutral expression. However, besides singer identity, emotional expression is also essential to convey the singer’s emotions and attitudes, but current SVC systems can not effectively support it. In this paper, we propose an expressive SVC framework called ESVC, which can convert singer identity and emotional style simultaneously. ESVC combines the ideas of style fusion and feature disentanglement, seeking to maximize fidelity in terms of emotional style and singer identity. Firstly, for style information penetration, we employ adaptive instance normalization (AdaIN) to fuse the content feature and style feature. Secondly, given the possibility of information leakage, two disentanglement-oriented methods are introduced to decouple different kinds of singing features. Mutual information (MI) is used to reduce the correlation between linguistic content, fundamental frequency (F0) and expressive feature, while adversarial triplet loss is exerted for decoupling identity and emotional elements. To the best of our knowledge, ESVC is the first SVC system to jointly convert singer identity and emotional style. Objective and subjective experiments demonstrate that our system significantly outperforms the state-of-the-art SVC model in terms of style expressiveness.

In recent years, the field of talking faces generation has attracted considerable attention, with certain methods adept at generating virtual faces that convincingly imitate human expressions. However, existing methods face challenges related to limited generalization, particularly when dealing with challenging identities. Furthermore, methods for editing expressions are often confined to a singular emotion, failing to adapt to intricate emotions. To overcome these challenges, this paper proposes EmoTalker, an emotionally editable portraits animation approach based on the diffusion model. EmoTalker modifies the denoising process to ensure preservation of the original portrait’s identity during inference. To enhance emotion comprehension from text input, Emotion Intensity Block is introduced to analyze fine-grained emotions and strengths derived from prompts. Additionally, a crafted dataset is harnessed to enhance emotion comprehension within prompts. Experiments show the effectiveness of EmoTalker in generating high-quality, emotionally customizable facial expressions.

Large Language Models (LLMs) have shown significant promise in various applications, including zero-shot and few-shot learning. However, their performance can be hampered by inherent biases. Instead of traditionally sought methods that aim to minimize or correct these biases, this study introduces a novel methodology named "bias-kNN". This approach capitalizes on the biased outputs, harnessing them as primary features for kNN and supplementing with gold labels. Our comprehensive evaluations, spanning diverse domain text classification datasets and different GPT-2 model sizes, indicate the adaptability and efficacy of the "bias-kNN" method. Remarkably, this approach not only outperforms conventional in-context learning in few-shot scenarios but also demonstrates robustness across a spectrum of samples, templates and verbalizers. This study, therefore, presents a unique perspective on harnessing biases, transforming them into assets for enhanced model performance.

In automatic speech recognition (ASR) task, the output sequence should correspond to a linear transcription of the input sequence. Lots of works have been done to learn the monotonic alignment in end-to-end (E2E) ASR model, but their methods mainly focus on streaming propose and usually result in a decline in ASR performance. On the contrary, some studies have shown that for non-streaming attention-based models, monotonic alignment is beneficial to model performance. Based on this motivation, we propose the enhanced Gaussian Monotonic Alignment (e-GMA), which reduces the difficulty of learning monotonic alignment, and the reconstructed attention matrix leads to an improved accuracy in ASR tasks. Experiments on the LibriSpeech dataset demonstrate the effectiveness of the proposed approach. Comparing with a strong baseline obtained from WeNet, the proposed model yields 12.2% relative WER reduction on test-clean benchmark and 9.9% on test-other.

In this work, we present DFlow, a novel generative framework that combines Normalizing Flow (NF) with a Denoising AutoEncoder (DAE), for high-fidelity waveform generation. With a tactfully designed structure, DFlow seamlessly integrates the capabilities of both NF and DAE, resulting in a significantly improved performance compared to the standard NF models. Experimental results showcase DFlow’s superiority, achieving the highest MOS score among the existing methods on commonly used datasets and the fastest synthesis speed among all likelihood models. We further demonstrate the generalization ability of DFlow by generating high-quality out-of-distribution audio samples, such as singing and music audio. Additionally, we extend the model capacity of DFlow by scaling up both the model size and training set size. Our large-scale universal vocoder, DFlow-XL, achieves highly competitive performance against the best universal vocoder, BigVGAN.

Voice conversion is the task to transform voice characteristics of source speech while preserving content information. Nowadays, self-supervised representation learning models are increasingly utilized in content extraction. However, in these representations, a lot of hidden speaker information leads to timbre leakage while the prosodic information of hidden units lacks use. To address these issues, we propose a novel framework for expressive voice conversion called “SAVC” based on soft speech units from HuBert-soft. Taking soft speech units as input, we design an attribute encoder to extract content and prosody features respectively. Specifically, we first introduce statistic perturbation imposed by adversarial style augmentation to eliminate speaker information. Then the prosody is implicitly modeled on soft speech units with knowledge distillation. Experiment results show that the intelligibility and naturalness of converted speech outperform previous work.

Single-model systems often suffer from deficiencies in tasks such as speaker verification (SV) and image classification, relying heavily on partial prior knowledge during decision-making, resulting in suboptimal performance. Although multi-model fusion (MMF) can mitigate some of these issues, redundancy in learned representations may limits improvements. To this end, we propose an adversarial complementary representation learning (ACoRL) framework that enables newly trained models to avoid previously acquired knowledge, allowing each individual component model to learn maximally distinct, complementary representations. We make three detailed explanations of why this works and experimental results demonstrate that our method more efficiently improves performance compared to traditional MMF. Furthermore, attribution analysis validates the model trained under ACoRL acquires more complementary knowledge, highlighting the efficacy of our approach in enhancing efficiency and robustness across tasks.

One-shot voice conversion aims to change the timbre of any source speech to match that of the unseen target speaker with only one speech sample. Existing methods face difficulties in satisfactory speech representation disentanglement and suffer from sizable networks as some of them leverage numerous complex modules for disentanglement. In this paper, we propose a model named MAIN-VC to effectively disentangle via a concise neural network. The proposed model utilizes Siamese encoders to learn clean representations, further enhanced by the designed mutual information estimator. The Siamese structure and the newly designed convolution module contribute to the lightweight of our model while ensuring performance in diverse voice conversion tasks. The experimental results show that the proposed model achieves comparable subjective scores and exhibits improvements in objective metrics compared to existing methods in a one-shot voice conversion scenario.

Using unsupervised learning to disentangle speech into content, rhythm, pitch, and timbre for voice conversion has become a hot research topic. Existing works generally take into account disentangling speech components through human-crafted bottleneck features which can not achieve sufficient information disentangling, while pitch and rhythm may still be mixed together. There is a risk of information overlap in the disentangling process which results in less speech naturalness. To overcome such limits, we propose a two-stage model to disentangle speech representations in a self-supervised manner without a human-crafted bottleneck design, which uses the Mutual Information (MI) with the designed upper bound estimator (IFUB) to separate overlapping information between speech components. Moreover, we design a Joint Text-Guided Consistent (TGC) module to guide the extraction of speech content and eliminate timbre leakage issues. Experiments show that our model can achieve a better performance than the baseline, regarding disentanglement effectiveness, speech naturalness, and similarity. Audio samples can be found at https://largeaudiomodel.com/eadvc.

Extractive Question Answering (EQA) in Machine Reading Comprehension (MRC) often faces the challenge of dealing with semantically identical but format-variant inputs. Our work introduces a novel approach, called the "Query Latent Semantic Calibrator (QLSC)", designed as an auxiliary module for existing MRC models. We propose a unique scaling strategy to capture latent semantic center features of queries. These features are then seamlessly integrated into traditional query and passage embeddings using an attention mechanism. By deepening the comprehension of the semantic queries-passage relationship, our approach diminishes sensitivity to variations in text format and boosts the model’s capability in pinpointing accurate answers. Experimental results on robust Question-Answer datasets confirm that our approach effectively handles format-variant but semantically identical queries, highlighting the effectiveness and adaptability of our proposed method.

With the rapid advancements in artificial intelligence, the development of knowledgeable and personalized agents has become increasingly prevalent. However, the inherent variability in state variables and action spaces among personalized agents poses significant aggregation challenges for traditional federated learning algorithms. To tackle these challenges, we introduce the Federated Split Decision Transformer (FSDT), an innovative framework designed explicitly for AI agent decision tasks. The FSDT framework excels at navigating the intricacies of personalized agents by harnessing distributed data for training while preserving data privacy. It employs a two-stage training process, with local embedding and prediction models on client agents and a global transformer decoder model on the server. Our comprehensive evaluation using the benchmark D4RL dataset highlights the superior performance of our algorithm in federated split learning for personalized agents, coupled with significant reductions in communication and computational overhead compared to traditional centralized training approaches. The FSDT framework demonstrates strong potential for enabling efficient and privacy-preserving collaborative learning in applications such as autonomous driving decision systems. Our findings underscore the efficacy of the FSDT framework in effectively leveraging distributed offline reinforcement learning data to enable powerful multi-type agent decision systems.

Singing voice beautifying is a novel task that has application value in people’s daily life, aiming to correct the pitch of the singing voice and improve the expressiveness without changing the original timbre and content. Existing methods rely on paired data or only concentrate on the correction of pitch. However, professional songs and amateur songs from the same person are hard to obtain, and singing voice beautifying doesn’t only contain pitch correction but other aspects like emotion and rhythm. Since we propose a fast and high-fidelity singing voice beautifying system called ConTuner, a diffusion model combined with the modified condition to generate the beautified Mel-spectrogram, where the modified condition is composed of optimized pitch and expressiveness. For pitch correction, we establish a mapping relationship from MIDI, spectrum envelope to pitch. To make amateur singing more expressive, we propose the expressiveness enhancer in the latent space to convert amateur vocal tone to professional. ConTuner achieves a satisfactory beautification effect on both Mandarin and English songs. Ablation study demonstrates that the expressiveness enhancer and generator-based accelerate method in ConTuner are effective.

In recent years, Transformer networks have shown remarkable performance in speech recognition tasks. However, their deployment poses challenges due to high computational and storage resource requirements. To address this issue, a lightweight model called EfficientASR is proposed in this paper, aiming to enhance the versatility of Transformer models. EfficientASR employs two primary modules: Shared Residual Multi-Head Attention (SRMHA) and Chunk-Level Feedforward Networks (CFFN). The SRMHA module effectively reduces redundant computations in the network, while the CFFN module captures spatial knowledge and reduces the number of parameters. The effectiveness of the EfficientASR model is validated on two public datasets, namely Aishell-1 and HKUST. Experimental results demonstrate a 36% reduction in parameters compared to the baseline Transformer network, along with improvements of 0.3% and 0.2% in Character Error Rate (CER) on the Aishell-1 and HKUST datasets, respectively.

Surveillance cameras are ubiquitous nowadays and users’ increasing needs for accessing real-world information (e.g., finding abandoned luggage) have urged object queries in real-time videos. While recent real-time video query processing systems exhibit excellent performance, they lack utility in deployment in practice as they overlook some crucial aspects, including multi-camera exploration, resource contention, and content awareness. Motivated by these issues, we propose a framework Gecko, to provide resource-efficient and accurate real-time object queries of massive videos on edge devices. Gecko (i) obtains optimal models from the model zoo and assigns them to edge devices for executing current queries, (ii) optimizes resource usage of the edge cluster at runtime by dynamically adjusting the frame query interval of each video stream and forking/joining running models on edge devices, and (iii) improves accuracy in changing video scenes by fine-grained stream transfer and continuous learning of models. Our evaluation with real-world video streams and queries shows that Gecko achieves up to 2x more resource efficiency gains and increases overall query accuracy by at least 12% compared with prior work, further delivering excellent scalability for practical deployment.

With recent advances in speech synthesis including text-to-speech (TTS) and voice conversion (VC) systems enabling the generation of ultra-realistic audio deepfakes, there is growing concern about their potential misuse. However, most deepfake (DF) detection methods rely solely on the fuzzy knowledge learned by a single model, resulting in performance bottlenecks and transparency issues. Inspired by retrieval-augmented generation (RAG), we propose a retrieval-augmented detection (RAD) framework that augments test samples with similar retrieved samples for enhanced detection. We also extend the multi-fusion attentive classifier to integrate it with our proposed RAD framework. Extensive experiments show the superior performance of the proposed RAD framework over baseline methods, achieving state-of-the-art results on the ASVspoof 2021 DF set and competitive results on the 2019 and 2021 LA sets. Further sample analysis indicates that the retriever consistently retrieves samples mostly from the same speaker with acoustic characteristics highly consistent with the query audio, thereby improving detection performance.

Current methods for few-shot action recognition mainly fall into the metric learning framework following ProtoNet, which demonstrates the importance of prototypes. Although they achieve relatively good performance, the effect of multimodal information is ignored, e.g. label texts. In this work, we propose a novel MultimOdal PRototype-ENhanced Network (MORN), which uses the semantic information of label texts as multimodal information to enhance prototypes. A CLIP visual encoder and a frozen CLIP text encoder are introduced to obtain features with good multimodal initialization. Then in the visual flow, visual prototypes are computed by a visual prototype-computed module. In the text flow, a semantic-enhanced (SE) module and an inflating operation are used to obtain text prototypes. The final multimodal prototypes are then computed by a multimodal prototype-enhanced (MPE) module. Besides, we define a PRototype SImilarity DiffErence (PRIDE) to evaluate the quality of prototypes, which is used to verify our improvement on the prototype level and effectiveness of MORN. We conduct extensive experiments on four popular few-shot action recognition datasets: HMDB51, UCF101, Kinetics and SSv2, and MORN achieves state-of-the-art results. When plugging PRIDE into the training stage, the performance can be further improved.

Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, Jing Xiao. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Recently multimodal medical pretraining models play a significant role in automatic medical image and text analysis that has wide social and economical impact in healthcare. Despite being able to be quickly transferred to downstream tasks, the models are greatly limited due to the fact that these models can only be pretrained with professional medical image-text datasets, which usually contain a very small number of samples. In this work We propose MITER (Medical Image-Text Joint adaptive Pretraining), a joint adaptive pretraining framework via multi-level contrastive learning to overcome this limitation by pretraining image and text models for medical domain and utilizing existing models pretrained on generic data, which contain enormous number of samples. MITER features two types of objectives to solve the problem. The first type is uni-modal objectives that pretrain the models with medical images and text separately on uni-modal tasks. The other type is a cross-modal objective that pretrains jointly, allowing the models to influence each other on cross-modal tasks. We also introduce a strategy to dynamically select hard negative samples during the training process for better performance. Experimental results over four medical tasks, image-report retrieval, multi-label image classification, visual question answering, and report generation, show that our MITER framework solves the limitation problem by greatly outperforming existing benchmark models on all the tasks. The source code of our framework is available online.2

Recently, the field of Text-to-Speech (TTS) has been dominated by one-stage text-to-waveform models which have significantly improved speech quality compared to two-stage models. In this work, we propose EfficientTTS 2 (EFTS2), a one-stage high-quality end-to-end TTS framework that is fully differentiable and highly efficient. Our method adopts an adversarial training process, with a differentiable aligner and a hierarchical-VAE-based waveform generator. These design choices free the model from the use of external aligners, invertible structures, and complex training procedures as most previous TTS works have. Moreover, we extend EFTS2 to the voice conversion (VC) task and propose EFTS2-VC, an end-to-end VC model that allows high-quality speech-to-speech conversion. Experimental results suggest that the two proposed models achieve better or at least comparable speech quality compared to baseline models, while also providing faster inference speeds and smaller model sizes.

Deep neural retrieval models have amply demonstrated their power but estimating the reliability of their predictions remains challenging. Most dialog response retrieval models output a single score for a response on how relevant it is to a given question. However, the bad calibration of deep neural network results in various uncertainty for the single score such that the unreliable predictions always misinform user decisions. To investigate these issues, we present an efficient calibration and uncertainty estimation framework PG-DRR for dialog response retrieval models which adds a Gaussian Process layer to a deterministic deep neural network and recovers conjugacy for tractable posterior inference by Pólya-Gamma augmentation. Finally, PG-DRR achieves the lowest empirical calibration error (ECE) in the in-domain datasets and the distributional shift task while keeping R10@1 and MAP performance.

Speaker verification (SV) performance deteriorates as utterances become shorter. To this end, we propose a new architecture called VoiceExtender which provides a promising solution for improving SV performance when handling short-duration speech signals. We use two guided diffusion models, the built-in and the external speaker embedding (SE) guided diffusion model, both of which utilize a diffusion model-based sample generator that leverages SE guidance to augment the speech features based on a short utterance. Extensive experimental results on the VoxCeleb1 dataset show that our method outperforms the baseline, with relative improvements in equal error rate (EER) of $46.1 \%, 35.7 \%$, $10.4 \%$, and $5.7 \%$ for the short utterance conditions of 0.5, $1.0,1.5$, and 2.0 seconds, respectively.

This paper proposes Shoggoth, an efficient edge-cloud collaborative architecture, for boosting inference performance on real-time video of changing scenes. Shoggoth uses online knowledge distillation to improve the accuracy of models suffering from data drift and offloads the labeling process to the cloud, alleviating constrained resources of edge devices. At the edge, we design adaptive training using small batches to adapt models under limited computing power, and adaptive sampling of training frames for robustness and reducing bandwidth. The evaluations on the realistic dataset show 15%–20% model accuracy improvement compared to the edge-only strategy and fewer network costs than the cloud-only strategy.

Zijian Zhang, Chang Shu, Ya Xiao, Yuan Shen, Di Zhu, Youxin Chen, Jing Xiao, Jey Han Lau, Qian Zhang, Zheng Lu. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 2023.

The Retrieval Question Answering (ReQA) task employs the retrieval-augmented framework, composed of a retriever and generator.The generator formulates the answer based on the documents retrieved by the retriever.Incorporating Large Language Models (LLMs) as generators is beneficial due to their advanced QA capabilities, but they are typically too large to be fine-tuned with budget constraints while some of them are only accessible via APIs.To tackle this issue and further improve ReQA performance, we propose a trainable Pluggable Reward-Driven Contextual Adapter (PRCA), keeping the generator as a black box.Positioned between the retriever and generator in a Pluggable manner, PRCA refines the retrieved information by operating in a tokenautoregressive strategy via maximizing rewards of the reinforcement learning phase.Our experiments validate PRCA's effectiveness in enhancing ReQA performance on three datasets by up to 20% improvement to fit black-box LLMs into existing frameworks, demonstrating its considerable potential in the LLMs era.

Out-of-distribution (OOD) detection aims at enhancing standard deep neural networks to distinguish anomalous inputs from original training data. Previous progress has introduced various approaches where the in-distribution training data and even several OOD examples are prerequisites. However, due to privacy and security, auxiliary data tends to be impractical in a real-world scenario. In this paper, we propose a data-free method without training on natural data, called Class-Conditional Impressions Reappearing (C2IR), which utilizes image impressions from the fixed model to recover class-conditional feature statistics. Based on that, we introduce Integral Probability Metrics to estimate layer-wise class-conditional deviations and obtain layer weights by Measuring Gradient-based Importance (MGI). The experiments verify the effectiveness of our method and indicate that C2IR outperforms other post-hoc methods and reaches comparable performance to the full access (ID and OOD) detection method, especially in the far-OOD dataset (SVHN).

Data-Free Knowledge Distillation (DFKD) has recently attracted growing attention in the academic community, especially with major breakthroughs in computer vision. Despite promising results, the technique has not been well applied to audio and signal processing. Due to the variable duration of audio signals, it has its own unique way of modeling. In this work, we propose feature-rich audio model inversion (FRAMI), a data-free knowledge distillation framework for general sound classification tasks. It first generates high-quality and feature-rich Mel-spectrograms through a feature-invariant contrastive loss. Then, the hidden states before and after the statistics pooling layer are reused when knowledge distillation is performed on these feature-rich samples. Experimental results on the Urbansound8k, ESC-50, and audioMNIST datasets demonstrate that FRAMI can generate feature-rich samples. Meanwhile, the accuracy of the student model is further improved by reusing the hidden state and significantly outperforms the baseline method.

Music genre classification has been widely studied in past few years for its various applications in music information retrieval. Previous works tend to perform unsatisfactorily, since those methods only use audio content or jointly use audio content and lyrics content inefficiently. In addition, as genres normally co-occur in a music track, it is desirable to capture and model the genre correlations to improve the performance of multi-label music genre classification. To solve these issues, we present a novel multi-modal method leveraging audio-lyrics contrastive loss and two symmetric cross-modal attention, to align and fuse features from audio and lyrics. Furthermore, based on the nature of the multi-label classification, a genre correlations extraction module is presented to capture and model potential genre correlations. Extensive experiments demonstrate that our proposed method significantly surpasses other multi-label music genre classification methods and achieves state-of-the-art result on Music4All dataset.

Non-parallel many-to-many voice conversion is a kind of style transfer task in speech. Recently, AutoVC has been applied in this field as a popular solution, as it can achieve distribution-matching style transfer by training only the re- construction loss. However, in order to strike a good balance between timbre disentanglement and sound quality, AutoVC requires imposing very strict constraints on the dimensionality of the latent representation. This constraint affects the quality of the converted speech while making it challenging to apply to other datasets directly. This paper proposes a new voice conversion framework that uses only one encoder to obtain timbre and content information by partitioning the latent space in the channel dimension. Furthermore, two different types of classifiers and two additional reconstruction losses are proposed to ensure that different parts of the latent space contain only separated content and timbre information, respectively. Experiments on the VCTK dataset show that the proposed model achieves state-of-the-art results in terms of the naturalness and similarity of converted speech. In addition, we experimentally show that for different division proportions of latent space, the content and timbre information will always be well separated.

Voice Conversion(VC) refers to converting the voice characteristics of audio to another one as it is said by other people. Recently, more and more studies have focused on disentangle-based VC, which separates the timbre and linguistic content information from an audio signal to effectively achieve VC tasks. However, It’s still challenging to extract phoneme-level features from frame-level hidden representations. This paper proposed a novel zero-shot voice conversion framework that utilizes contrastive learning and vector quantization to encourage the frame-level hidden features closer to the phoneme-level linguistic information, called VQ-CL. All objective and subjective experiment results show that VQ-CL has better performance than previous studies in separating content and voice characteristics to improve the sound quality of generated speech.

Recent expressive text to speech (TTS) models focus on synthesizing emotional speech, but some fine-grained styles such as intonation are neglected. In this paper, we propose QI-TTS which aims to better transfer and control intonation to further deliver the speaker’s questioning intention while transferring emotion from reference speech. We propose a multi-style extractor to extract style embedding from two different levels. While the sentence level represents emotion, the final syllable level represents intonation. For fine-grained intonation control, we use relative attributes to represent intonation intensity at the syllable level. Experiments have validated the effectiveness of QI-TTS for improving intonation expressiveness in emotional speech synthesis.

Deep neural networks have achieved remarkable performance in retrieval-based dialogue systems, but they are shown to be ill calibrated. Though basic calibration methods like Monte Carlo Dropout and Ensemble can calibrate well, these methods are time-consuming in the training or inference stages. To tackle these challenges, we propose an efficient uncertainty calibration framework GPF-BERT for BERT-based conversational search, which employs a Gaussian Process layer and the focal loss on top of the BERT architecture to achieve a high-quality neural ranker. Extensive experiments are conducted to verify the effectiveness of our method. In comparison with basic calibration methods, GPF-BERT achieves the lowest empirical calibration error (ECE) in three in-domain datasets and the distributional shift tasks, while yielding the highest R10@1 and MAP performance on most cases. In terms of time consumption, our GPF-BERT has an 8× speedup.

Because of predicting all the target tokens in parallel, the non-autoregressive models greatly improve the decoding efficiency of speech recognition compared with traditional autoregressive models. In this work, we present dynamic alignment Mask CTC, introducing two methods: (1) Aligned Cross Entropy (AXE), finding the monotonic alignment that minimizes the cross-entropy loss through dynamic programming, (2) Dynamic Rectification, creating new training samples by replacing some masks with model predicted tokens. The AXE ignores the absolute position alignment between prediction and ground truth sentence and focuses on tokens matching in relative order. The dynamic rectification method makes the model capable of simulating the non-mask but possible wrong tokens, even if they have high confidence. Our experiments on WSJ dataset demonstrated that not only AXE loss but also the rectification method could improve the WER performance of Mask CTC.

Using deep learning methods to classify EEG signals can accurately identify people’s emotions. However, existing studies have rarely considered the application of the information in another domain’s representations to feature selection in the time-frequency domain. We propose a classification network of EEG signals based on the cross-domain feature fusion method, which makes the network more focused on the features most related to brain activities and thinking changes by using the multi-domain attention mechanism. In addition, we propose a two-step fusion method and apply these methods to the EEG emotion recognition network. Experimental results show that our proposed network, which combines multiple representations in the time-frequency domain and spatial domain, outperforms previous methods on public datasets and achieves state-of-the-art at present.

Cross-modal retrieval (CMR) has been extensively applied in various domains, such as multimedia search engines and recommendation systems. Most existing CMR methods focus on image-to-text retrieval, whereas audio-to-text retrieval, a less explored domain, has posed a great challenge due to the difficulty to uncover discriminative features from audio clips and texts. Existing studies are restricted in the following two ways: 1) Most researchers utilize contrastive learning to construct a common subspace where similarities among data can be measured. However, they considers only cross-modal transformation, neglecting the intra-modal separability. Besides, the temperature parameter is not adaptively adjusted along with semantic guidance, which degrades the performance. 2) These methods do not take latent representation reconstruction into account, which is essential for semantic alignment. This paper introduces a novel audio-text oriented CMR approach, termed Contrastive Latent Space Reconstruction Learning (CLSR). CLSR improves contrastive representation learning by taking intra-modal separability into account and adopting an adaptive temperature control strategy. Moreover, the latent representation reconstruction modules are embedded into the CMR framework, which improves modal interaction. Experiments in comparison with some state-of-the-art methods on two audio-text datasets have validated the superiority of CLSR.

Most existing sandstorm image enhancement methods are based on traditional theory and prior knowledge, which often restrict their applicability in real-world scenarios. In addition, these approaches often adopt a strategy of color correction followed by dust removal, which makes the algorithm structure too complex. To solve the issue, we introduce a novel image restoration model, named all-in-one sandstorm removal network (AOSR-Net). This model is developed based on a re-formulated sandstorm scattering model, which directly establishes the image mapping relationship by integrating intermediate parameters. Such integration scheme effectively addresses the problems of over-enhancement and weak generalization in the field of sand dust image enhancement. Experimental results on synthetic and real-world sandstorm images demonstrate the superiority of the proposed AOSR-Net over state-of-the-art (SOTA) algorithms.

This paper integrates graph-to-sequence into an end-to-end text-to-speech framework for syntax-aware modelling with syntactic information of input text. Specifically, the input text is parsed by a dependency parsing module to form a syntactic graph. The syntactic graph is then encoded by a graph encoder to extract the syntactic hidden information, which is concatenated with phoneme embedding and input to the alignment and flow-based decoding modules to generate the raw audio waveform. The model is experimented on two languages, English and Mandarin, using single-speaker, few samples of target speakers, and multi-speaker datasets, respectively. Experimental results show better prosodic consistency performance between input text and generated audio, and also get higher scores in the subjective prosodic evaluation, and show the ability of voice conversion. Besides, the efficiency of the model is largely boosted through the design of the AI chip operator with 5x acceleration.

Federated Learning (FL) has been widely concerned for it enables decentralized learning while ensuring data privacy. However, most existing methods unrealistically assume that the classes encountered by local clients are fixed over time. After learning new classes, this impractical assumption will make the model's catastrophic forgetting of old classes significantly severe. Moreover, due to the limitation of communication cost, it is challenging to use large-scale models in FL, which will affect the prediction accuracy. To address these challenges, we propose a novel framework, Federated Enhanced Transformer (FedET), which simultaneously achieves high accuracy and low communication cost. Specifically, FedET uses Enhancer, a tiny module, to absorb and communicate new knowledge, and applies pre-trained Transformers combined with different Enhancers to ensure high precision on various tasks. To address local forgetting caused by new classes of new tasks and global forgetting brought by non-i.i.d class imbalance across different local clients, we proposed an Enhancer distillation method to modify the imbalance between old and new knowledge and repair the non-i.i.d. problem. Experimental results demonstrate that FedET's average accuracy on a representative benchmark dataset is 14.1% higher than the state-of-the-art method, while FedET saves 90% of the communication cost compared to the previous method.

Text summarization is essential for information aggregation and demands large amounts of training data. However, concerns about data privacy and security limit data collection and model training. To eliminate this concern, we propose a federated learning text summarization scheme, which allows users to share the global model in a cooperative learning manner without sharing raw data. Personalized federated learning (PFL) balances personalization and generalization in the process of optimizing the global model, to guide the training of local models. However, multiple local data have different distributions of semantics and context, which may cause the local model to learn deviated semantic and context information. In this paper, we propose FedSUMM, a dynamic gradient adapter to provide more appropriate local parameters for local model. Simultaneously, FedSUMM uses differential privacy to prevent parameter leakage during distributed training. Experimental evidence verifies FedSUMM can achieve faster model convergence on PFL algorithm for task-specific text summarization, and the method achieves superior performance for different optimization metrics for text summarization.

In recent Text-to-Speech (TTS) systems, a neural vocoder often generates speech samples by solely conditioning on acoustic features predicted from an acoustic model. However, there are always distortions existing in the predicted acoustic features, compared to those of the groundtruth, especially in the common case of poor acoustic modeling due to low-quality training data. To overcome such limits, we propose a Self-supervised learning framework to learn an Anti-distortion acoustic Representation (SAR) to replace human-crafted acoustic features by introducing distortion prior to an auto-encoder pre-training process. The learned acoustic representation from the proposed framework is proved anti-distortion compared to the most commonly used mel-spectrogram through both objective and subjective evaluation.

Conversational Question Answering (CQA) is a challenging task that aims to generate natural answers for conversational flow questions.In this paper, we propose a pluggable approach for extractive methods that introduces a novel prompt-guided copy mechanism to improve the fluency and appropriateness of the extracted answers.Our approach uses prompts to link questions to answers and employs attention to guide the copy mechanism to verify the naturalness of extracted answers, making necessary edits to ensure that the answers are fluent and appropriate.The three prompts, including a question-rationale relationship prompt, a question description prompt, and a conversation history prompt, enhance the copy mechanism's performance.Our experiments demonstrate that this approach effectively promotes the generation of natural answers and achieves good results in the CoQA challenge.

Voice conversion as the style transfer task applied to speech, refers to converting one person's speech into a new speech that sounds like another person's. Up to now, there has been a lot of research devoted to better implementation of VC tasks. However, a good voice conversion model should not only match the timbre information of the target speaker, but also expressive information such as prosody, pace, pause, etc. In this context, prosody modeling is crucial for achieving expressive voice conversion that sounds natural and convincing. Unfortunately, prosody modeling is important but challenging, especially without text transcriptions. In this paper, we firstly propose a novel voice conversion framework named 'PMVC', which effectively separates and models the content, timbre, and prosodic information from the speech without text transcriptions. Specially, we introduce a new speech augmentation algorithm for robust prosody extraction. And building upon this, mask and predict mechanism is applied in the disentanglement of prosody and content information. The experimental results on the AIShell-3 corpus supports our improvement of naturalness and similarity of converted speech.

Detecting out-of-distribution (OOD) examples is crucial to guarantee the reliability and safety of deep neural networks in real-world settings. In this paper, we offer an innovative perspective on quantifying the disparities between in-distribution (ID) and OOD data---analyzing the uncertainty that arises when models attempt to explain their predictive decisions. This perspective is motivated by our observation that gradient-based attribution methods encounter challenges in assigning feature importance to OOD data, thereby yielding divergent explanation patterns. Consequently, we investigate how attribution gradients lead to uncertain explanation outcomes and introduce two forms of abnormalities for OOD detection: the zero-deflation abnormality and the channel-wise average abnormality. We then propose GAIA, a simple and effective approach that incorporates Gradient Abnormality Inspection and Aggregation. The effectiveness of GAIA is validated on both commonly utilized (CIFAR) and large-scale (ImageNet-1k) benchmarks. Specifically, GAIA reduces the average FPR95 by 23.10% on CIFAR10 and by 45.41% on CIFAR100 compared to advanced post-hoc methods.

Blind Omnidirectional Image Quality Assessment (BOIQA) aims to objectively assess the human perceptual quality of omnidirectional images (ODIs) without relying on pristine-quality image information. It is becoming more significant with the increasing advancement of virtual reality (VR) technology. However, the quality assessment of ODIs is severely hampered by the fact that the existing BOIQA pipeline lacks the modeling of the observer's browsing process. To tackle this issue, we propose a novel multi-sequence network for BOIQA called Assessor360, which is derived from the realistic multi-assessor ODI quality assessment procedure. Specifically, we propose a generalized Recursive Probability Sampling (RPS) method for the BOIQA task, combining content and details information to generate multiple pseudo viewport sequences from a given starting point. Additionally, we design a Multi-scale Feature Aggregation (MFA) module with a Distortion-aware Block (DAB) to fuse distorted and semantic features of each viewport. We also devise Temporal Modeling Module (TMM) to learn the viewport transition in the temporal domain. Extensive experimental results demonstrate that Assessor360 outperforms state-of-the-art methods on multiple OIQA datasets. The code and models are available at https://github.com/TianheWu/Assessor360.

Osteoporosis is a common chronic metabolic bone disease often under-diagnosed and under-treated due to the limited access to bone mineral density (BMD) examinations, e.g., via Dual-energy X-ray Absorptiometry (DXA). This paper proposes a method to predict BMD from Chest X-ray (CXR), one of the most commonly accessible and low-cost medical imaging examinations. The proposed method first automatically detects Regions of Interest (ROIs) of local CXR bone structures. Then a multi-ROI deep model with transformer encoder is developed to exploit both local and global information in the chest X-ray image for accurate BMD estimation. The proposed method is evaluated on 13719 CXR patient cases with ground truth BMD measured by the gold standard DXA. The model predicted BMD has a strong correlation with the ground truth (Pearson correlation coefficient 0.894 on lumbar 1). When applied in osteoporosis screening, it achieves a high classification performance (average AUC of 0.968). As the first effort of using CXR scans to predict the BMD, the proposed algorithm holds strong potential to promote early osteoporosis screening and public health.

Deep learning methods can struggle to handle domain shifts not seen in training data, which can cause them to not generalize well to unseen domains. This has led to research attention on domain generalization (DG), which aims to the model's generalization ability to out-of-distribution. Adversarial domain generalization is a popular approach to DG, but conventional approaches (1) struggle to sufficiently align features so that local neighborhoods are mixed across domains; and (2) can suffer from feature space over collapse which can threaten generalization performance. To address these limitations, we propose localized adversarial domain generalization with space compactness maintenance (LADG) which constitutes two major contributions. First, we propose an adversarial localized classifier as the domain discriminator, along with a principled primary branch. This constructs a min-max game whereby the aim of the featurizer is to produce locally mixed domains. Second, we propose to use a coding-rate loss to alleviate feature space over collapse. We conduct comprehensive experiments on the Wilds DG benchmark to validate our approach, where LADG outperforms leading competitors on most datasets.

Machine learning models (mainly neural networks) are used more and more in real life. Users feed their data to the model for training. But these processes are often one-way. Once trained, the model remembers the data. Even when data is removed from the dataset, the effects of these data persist in the model. With more and more laws and regulations around the world protecting data privacy, it becomes even more important to make models forget this data completely through machine unlearning.This paper adopts the projection residual method based on Newton iteration method. The main purpose is to implement machine unlearning tasks in the context of linear regression models and neural network models. This method mainly uses the iterative weighting method to completely forget the data and its corresponding influence, and its computational cost is linear in the feature dimension of the data. This method can improve the current machine learning method. At the same time, it is independent of the size of the training set. Results were evaluated by feature injection testing (FIT). Experiments show that this method is more thorough in deleting data, which is close to model retraining.

The extraction of sequence patterns from a collection of functionally linked unlabeled DNA sequences is known as DNA motif discovery, and it is a key task in computational biology. Several deep learning-based techniques have recently been introduced to address this issue. However, these algorithms can not be used in real-world situations because of the need for labeled data. Here, we presented RL-MD, a novel reinforcement learning based approach for DNA motif discovery task. RL-MD takes unlabelled data as input, employs a relative information-based method to evaluate each proposed motif, and utilizes these continuous evaluation results as the reward. The experiments show that RL-MD can identify high-quality motifs in real-world data.

The Transformer architecture model, based on self-attention and multi-head attention, has achieved remarkable success in offline end-to-end Automatic Speech Recognition (ASR). However, self-attention and multi-head attention cannot be easily applied for streaming or online ASR. For self-attention in Transformer ASR, the softmax normalization function-based attention mechanism makes it impossible to highlight important speech information. For multi-head attention in Transformer ASR, it is not easy to model monotonic alignments in different heads. To overcome these two limits, we integrate sparse attention and monotonic attention into Transformer-based ASR. The sparse mechanism introduces a learned sparsity scheme to enable each self-attention structure to fit the corresponding head better. The monotonic attention deploys regularization to prune redundant heads for the multi-head attention structure. The experiments show that our method can effectively improve the attention mechanism on widely used benchmarks of speech recognition.

Machine Learning as a Service is a promising service for individuals and companies who would like to delegate model training to third parties. The customers desire proof of the integrity of the model training to prevent potential backdoor attacks launched by the server, while the server desires to prove the integrity without revealing their intellectual assets, hyper-parameters of the training scheme. Zero-knowledge proof, a cryptographic tool can theoretically satisfy the above demand, but is still practically infeasible due to the inefficiency of proving. Thus, we propose zkMLaaS, a privacy-preserving and verifiable scheme for efficient training proof generation in the MLaaS scenario. zkMLaaS features a two-round challenge-response pro-tocol equipped with the random sampling. This greatly reduces the time cost of proof generation and ensures the integrity of training procedure simultaneously. We analyze the security of zkMLaaS and conduct comprehensive evaluation which shows it saves around$273\times$times compared with naive scheme.

Existing methods for speaker age estimation usually treat it as a multi-class classification or a regression problem. However, precise age identification remains a challenge due to label ambiguity, i.e., utterances from adjacent age of the same person are often indistinguishable. To address this, we utilize the ambiguous information among the age labels, convert each age label into a discrete label distribution and leverage the label distribution learning (LDL) method to fit the data. For each audio data sample, our method produces a age distribution of its speaker, and on top of the distribution we also perform two other tasks: age prediction and age uncertainty minimization. Therefore, our method naturally combines the age classification and regression approaches, which enhances the robustness of our method. We conduct experiments on the public NIST SRE08-10 dataset and a real-world dataset, which exhibit that our method outperforms baseline methods by a relatively large margin, yielding a 10% reduction in terms of mean absolute error (MAE) on a real-world dataset.

Voice Conversion(VC) refers to changing the timbre of a speech while retaining the discourse content. Recently, many works have focused on disentangle-based learning techniques to separate the timbre and the linguistic content information from a speech signal. Once successful, voice conversion will be feasible and straightforward. This paper proposed a novel one-shot voice conversion framework based on vector quantization voice conversion (VQVC) and AutoVC, called AVQVC. A new training method is applied to VQVC to separate content and timbre information from speech more effectively. The result shows that this approach has better performance than VQVC in separating content and timbre to improve the sound quality of generated speech.

Any-to-any voice conversion problem aims to convert voices for source and target speakers, which are out of the training data. Previous works wildly utilize the disentangle-based models. The disentangle-based model assumes the speech consists of content and speaker style information and aims to untangle them to change the style information for conversion. Previous works focus on reducing the dimension of speech to get the content information. But the size is hard to determine to lead to the untangle overlapping problem. We propose the Disentangled Representation Voice Conversion (DRVC) model to address the issue. DRVC model is an end-to-end self-supervised model consisting of the content encoder, timbre encoder, and generator. Instead of the previous work for reducing speech size to get content, we propose a cycle for restricting the disentanglement by the Cycle Reconstruct Loss and Same Loss. The experiments show there is an improvement for converted speech on quality and voice similarity.

The visual dialog task attempts to train an agent to answer multi-turn questions given an image, which requires the deep understanding of interactions between the image and dialog history. Existing researches tend to employ the modality-specific modules to model the interactions, which might be troublesome to use. To fill in this gap, we propose a unified framework for image-text joint embedding, named VU-BERT, and apply patch projection to obtain vision embedding firstly in visual dialog tasks to simplify the model. The model is trained over two tasks: masked language modeling and next utterance retrieval. These tasks help in learning visual concepts, utterances dependence, and the relationships between these two modalities. Finally, our VU-BERT achieves competitive performance (0.7287 NDCG scores) on VisDial v1.0 Datasets.

Incomplete utterance rewriting (IUR) has recently become an essential task in NLP, aiming to complement the incomplete utterance with sufficient context information for comprehension. In this paper, we propose a novel method by directly extracting the coreference and omission relationship from the self-attention weight matrix of the transformer in-stead of word embeddings and edit the original text accordingly to generate the complete utterance. Benefiting from the rich information in the self-attention weight matrix, our method achieved competitive results on public IUR datasets.

Grapheme-to-phoneme (G2P) conversion is the process of converting the written form of words to their pronunciations. It has an important role for text-to-speech (TTS) synthesis and automatic speech recognition (ASR) systems. In this paper, we aim to evaluate and enhance the robustness of G2P models. We show that neural G2P models are extremely sensitive to orthographical variations in graphemes like spelling mistakes. To solve this problem, we propose three controlled noise introducing methods to synthesize noisy training data. Moreover, we incorporate the contextual information with the baseline and propose a robust training strategy to stabilize the training process. The experimental results demonstrate that our proposed robust G2P model (r-G2P) outperforms the baseline significantly (-2.73% WER on Dict-based benchmarks and -9.09% WER on Real-world sources).

Multi-speaker text-to-speech (TTS) using a few adaption data is a challenge in practical applications. To address that, we propose a zero-shot multi-speaker TTS, named nnSpeech, that could synthesis a new speaker voice without fine-tuning and using only one adaption utterance. Compared with using a speaker representation module to extract the characteristics of new speakers, our method bases on a speaker-guided conditional variational autoencoder and can generate a variable Z, which contains both speaker characteristics and content information. The latent variable Z distribution is approximated by another variable conditioned on reference mel-spectrogram and phoneme. Experiments on the English corpus, Mandarin corpus, and cross-dataset proves that our model could generate natural and similar speech with only one adaption speech.

Buddhism is an influential religion with a long-standing history and profound philosophy. Nowadays, more and more people worldwide aspire to learn the essence of Buddhism, attaching importance to Buddhism dissemination. However, Buddhist scriptures written in classical Chinese are obscure to most people and machine translation applications. For instance, general Chinese-English neural machine translation (NMT) fails in this domain. In this paper, we proposed a novel approach to building a practical NMT model for Buddhist scriptures. The performance of our translation pipeline acquired highly promising results in ablation experiments under three criteria.

Since the beginning of the COVID-19 pandemic, remote conferencing and school-teaching have become important tools. The previous applications aim to save the commuting cost with real-time interactions. However, our application is going to lower the production and reproduction costs when preparing the communication materials. This paper proposes a system called Pre-Avatar, generating a presentation video with a talking face of a target speaker with 1 front-face photo and a 3-minute voice recording. Technically, the system consists of three main modules, user experience interface (UEI), talking face module and few-shot text-to-speech (TTS) module. The system firstly clones the target speaker's voice, and then generates the speech, and finally generate an avatar with appropriate lip and head movements. Under any scenario, users only need to replace slides with different notes to generate another new video. The demo has been released here11https://pre-avatar.github.io/ and will be published as free software for use.

Machine learning techniques have been widely applied in modern financial activities. Participants in the field are aware of the importance of data privacy. Vertical federated learning (VFL) was proposed as a solution to multi-party secure computation for machine learning to obtain the huge data required by the models as well as keep the privacy of the data holders. However, previous research majorly analyzed the algorithms under ideal conditions. Data imbalance in VFL is still an open problem. In this paper, we propose a privacy-preserving sampling strategy for imbalanced VFL based on federated graph embedding of the samples, without leaking any distribution information. The participants of the federation provide partial neighbor information for each sample during the intersection stage and the controversial negative sample will be filtered out. Experiments were conducted on commonly used financial datasets and one real-world dataset. Our proposed approach obtained the leading F1 score on all tested datasets on comparing with the baseline under sampling strategies for VFL.

Quantum devices with low qubits are common in the Noisy Intermediate-Scale Quantum (NISQ) era. However, Quantum Neural Network (QNN) running on low-qubit quantum devices would be difficult since it is based on Variational Quantum Circuit (VQC), which requires many qubits. Therefore, it is critical to make QNN with VQC run on low-qubit quantum devices. In this study, we propose a novel VQC called the low-qubit VQC. VQC requires numerous qubits based on the input dimension; however, the low-qubit VQC with linear transformation can liberate this condition. Thus, it allows the QNN to run on low-qubit quantum devices for speech applications. Furthermore, as compared to the VQC, our proposed low-qubit VQC can stabilize the training process more. Based on the low-qubit VQC, we implement QSpeech11Our implementation of QSpeech is publicly available at https://github.com/zhenhouhong/QSpeech, a library for quick prototyping of hybrid quantum-classical neural networks in the speech field. It has numerous quantum neural layers and QNN models for speech applications. Experiments on Speech Command Recognition and Text-to-Speech show that our proposed low-qubit VQC outperforms VQC and is more stable.

Pre-trained BERT models have achieved impressive performance in many natural language processing (NLP) tasks. However, in many real-world situations, textual data are usually decentralized over many clients and unable to be uploaded to a central server due to privacy protection and regulations. Federated learning (FL) enables multiple clients collaboratively to train a global model while keeping the local data privacy. A few researches have investigated BERT in federated learning setting, but the problem of performance loss caused by heterogeneous (e.g., non-IID) data over clients remain under-explored. To address this issue, we propose a framework, FedSplitBERT, which handles heterogeneous data and decreases the communication cost by splitting the BERT encoder layers into local part and global part. The local part parameters are trained by the local client only while the global part parameters are trained by aggregating gradients of multiple clients. Due to the sheer size of BERT, we explore a quantization method to further reduce the communication cost with minimal performance loss. Our framework is ready-to-use and compatible to many existing federated learning algorithms, including FedAvg, FedProx and FedAdam. Our experiments verify the effectiveness of the proposed framework, which outperforms baseline methods by a significant margin, while FedSplitBERT with quantization can reduce the communication cost by 11.9×.

In this paper, we investigated a speech augmentation based unsupervised learning approach for keyword spotting (KWS) task. KWS is a useful speech application, yet also heavily depends on the labeled data. We designed a CNN-Attention architecture to conduct the KWS task. CNN layers focus on the local acoustic features, and attention layers model the long-time dependency. To improve the robustness of KWS model, we also proposed an unsupervised learning method. The unsupervised loss is based on the similarity between the original and augmented speech features, as well as the audio reconstructing information. Two speech augmentation methods are explored in the unsupervised learning: speed and intensity. The experiments on Google Speech Commands V2 Dataset demonstrated that our CNN-Attention model has competitive results. Moreover, the augmentation based unsupervised learning could further improve the classification accuracy of KWS task. In our experiments, with augmentation based unsupervised learning, our KWS model achieves better performance than other unsupervised methods, such as CPC, APC, and MPC.

Low resource automatic speech recognition (ASR) is a useful but thorny task, since deep learning ASR models usually need huge amounts of training data. The existing models mostly established a bottleneck (BN) layer by pre-training on a large source language, and transferring to the low resource target language. In this work, we introduced an adaptive activation network to the upper layers of ASR model, and applied different activation functions to different languages. We also proposed two approaches to train the model: (1) cross-lingual learning, replacing the activation function from source language to target language, (2) multilingual learning, jointly training the Connectionist Temporal Classification (CTC) loss of each language and the relevance of different languages. Our experiments on IARPA Babel datasets demonstrated that our approaches outperform the from-scratch training and traditional bottleneck feature based methods. In addition, combining the cross-lingual learning and multilingual learning together could further improve the performance of multilingual speech recognition.

Non-negative matrix factorization (NMF) based topic modeling is widely used in natural language processing (NLP) to uncover hidden topics of short text documents. Usually, training a high-quality topic model requires large amount of textual data. In many real-world scenarios, customer textual data should be private and sensitive, precluding uploading to data centers. This paper proposes a Federated NMF (FedNMF) framework, which allows multiple clients to collaboratively train a high-quality NMF based topic model with locally stored data. However, standard federated learning will significantly undermine the performance of topic models in downstream tasks (e.g., text classification) when the data distribution over clients is heterogeneous. To alleviate this issue, we further propose FedNMF+MI, which simultaneously maximizes the mutual information (MI) between the count features of local texts and their topic weight vectors to mitigate the performance degradation. Experimental results show that our FedNMF+MI methods outperform Federated Latent Dirichlet Allocation (FedLDA) and the FedNMF without MI methods for short texts by a significant margin on both coherence score and classification F1 score.

Facial micro-expressions recognition has attracted much attention recently. Micro-expressions have the characteristics of short duration and low intensity, and it is difficult to train a high-performance classifier with the limited number of existing micro-expressions. Therefore, recognizing micro-expressions is a challenge task. In this paper, we propose a micro-expression recognition method based on attribute information embedding and cross-modal contrastive learning. We use 3D CNN to extract RGB features and FLOW features of micro-expression sequences and fuse them, and use BERT network to extract text information in Facial Action Coding System. Through cross-modal contrastive loss, we embed attribute information in the visual network, thereby improving the representation ability of micro-expression recognition in the case of limited samples. We conduct extensive experiments in CASME II and MMEW databases, and the accuracy is 77.82% and 71.04%, respectively. The comparative experiments show that this method has better recognition effect than other methods for microexnression recognition.

Federated learning (FL) is a paradigm where many clients collaboratively train a model under the coordination of a central server, while keeping the training data locally stored. However, heterogeneous data distributions over different clients remain a challenge to mainstream FL algorithms, which may cause slow convergence, overall performance degradation and unfairness of performance across clients. To address these problems, in this study we propose a reinforcement learning framework, called PG-FFL, which automatically learns a policy to assign aggregation weights to clients. Additionally, we propose to utilize Gini coefficient as the measure of fairness for FL. More importantly, we apply the Gini coefficient and validation accuracy of clients in each communication round to construct a reward function for the reinforcement learning. Our PG-FFL is also compatible to many existing FL algorithms. We conduct extensive experiments over diverse datasets to verify the effectiveness of our framework. The experimental results show that our framework can outperform baseline methods in terms of overall performance, fairness and convergence speed.

Deep learning models have made significant progress in automatic program repair. However, the black-box nature of these methods has restricted their practical applications. To address this challenge, this paper presents an interpretable approach for program repair based on sequence-to-sequence models with causal inference and our method is called CPR, short for causal program repair. Our CPR can generate explanations in the process of decision making, which consists of groups of causally related input-output tokens. Firstly, our method infers these relations by querying the model with inputs disturbed by data augmentation. Secondly, it generates a graph over tokens from the responses and solves a partitioning problem to select the most relevant components. The experiments on four programming languages (Java, C, Python, and JavaScript) show that CPR can generate causal graphs for reasonable interpretations and boost the performance of bug fixing in automatic program repair.

Deep neural networks have become popular in many supervised learning tasks, but they may suffer from overfitting when the training dataset is limited. To mitigate this, many researchers use data augmentation, which is a widely used and effective method for increasing the variety of datasets. However, the randomness introduced by data augmentation causes inevitable inconsistency between training and inference, which leads to poor improvement. In this paper, we propose a consistency regularization framework based on data augmentation, called CR-Aug, which forces the output distributions of different sub models generated by data augmentation to be consistent with each other. Specifically, CR-Aug evaluates the discrepancy between the output distributions of two augmented versions of each sample, and it utilizes a stop-gradient operation to minimize the consistency loss. We implement CR-Aug to image and audio classification tasks and conduct extensive experiments to verify its effectiveness in improving the generalization ability of classifiers. Our CR-Aug framework is ready-to-use, it can be easily adapted to many state-of-the-art network architectures. Our empirical results show that CR-Aug outperforms baseline methods by a significant margin.

With the rapid development of the Metaverse, virtual humans have emerged, and human image synthesis and editing techniques, such as pose transfer, have recently become popular. Most of the existing techniques rely on GANs, which can generate good human images even with large variants and occlusions. But from our best knowledge, the existing state-of-the-art method still has the following problems: the first is that the rendering effect of the synthetic image is not realistic, such as poor rendering of some regions. And the second is that the training of GAN is unstable and slow to converge, such as model collapse. Based on the above two problems, we propose several methods to solve them. To improve the rendering effect, we use the Residual Fast Fourier Transform Block to replace the traditional Residual Block. Then, spectral normalization and Wasserstein distance are used to improve the speed and stability of GAN training. Experiments demonstrate that the methods we offer are effective at solving the problems listed above, and we get state-of-the-art scores in LPIPS and PSNR.

Integrating multimodal knowledge for abstractive summarization task is a work-in-progress research area, with present techniques inheriting fusion-then-generation paradigm. Due to semantic gaps between computer vision and natural language processing, current methods often treat multiple data points as separate objects and rely on attention mechanisms to search for connection in order to fuse together. In addition, missing awareness of cross-modal matching from many frameworks leads to performance reduction. To solve these two drawbacks, we propose an Iterative Contrastive Alignment Framework (ICAF) that uses recurrent alignment and contrast to capture the coherences between images and texts. Specifically, we design a recurrent alignment (RA) layer to gradually investigate fine-grained semantical relationships between image patches and text tokens. At each step during the encoding process, crossmodal contrastive losses are applied to directly optimize the embedding space. According to ROUGE, relevance scores, and human evaluation, our model outperforms the state-of-the-art baselines on MSMO dataset. Experiments on the applicability of our proposed framework and hyperparameters settings have been also conducted.

Singing voice synthesis is a generative task that involves multi-dimensional control of the singing model, including lyrics, pitch, and duration, and includes the timbre of the singer and singing skills such as vibrato. In this paper, we proposed SU-net for singing voice synthesis named SUSing. Synthesizing singing voice is treated as a translation task between lyrics and music score and spectrum. The lyrics and music score information is encoded into a two-dimensional feature representation through the convolution layer. The two-dimensional feature and its frequency spectrum are mapped to the target spectrum in an autoregressive manner through a SU-net network. Within the SU-net the stripe pooling method is used to replace the alternate global pooling method to learn the vertical frequency relationship in the spectrum and the changes of frequency in the time domain. The experimental results on the public dataset Kiritan show that the proposed method can synthesize more natural singing voices.

Most singer identification methods are processed in the frequency domain, which potentially leads to information loss during the spectral transformation. In this paper, instead of the frequency domain, we propose an end-to-end architecture that addresses this problem in the waveform domain. An en-coder based on Multi-scale Dilated Convolution Neural Networks (MDCNN) was introduced to generate wave embedding from the raw audio signal. Specifically, dilated convolution layers are used in the proposed method to enlarge the receptive field, aiming to extract song-level features. Furthermore, skip connection in the backbone network integrates the multi-resolution acoustic features learned by the stack of convolution layers. Then, the obtained wave embedding is passed into the following networks for singer identification. In experiments, the proposed method achieves comparable performance on the benchmark dataset of Artist20, which significantly improves related works.

Recently, synthesizing personalized speech by text-to-speech (TTS) application is highly demanded. But the previous TTS models require a mass of target speaker speeches for training. It is a high-cost task, and hard to record lots of utterances from the target speaker. Data augmentation of the speeches is a solution but leads to the low-quality synthesis speech problem. Some multi-speaker TTS models are proposed to address the issue. But the quantity of utterances of each speaker imbalance leads to the voice similarity problem. We propose the Target Domain Adaptation Speech Synthesis Network (TDASS) to address these issues. Based on the backbone of the Tacotron2 model, which is the high-quality TTS model, TDASS introduces a self-interested classifier for reducing the non-target influence. Besides, a special gradient reversal layer with different operations for target and non-target is added to the classifier. We evaluate the model on a Chinese speech corpus, the experiments show the proposed method outperforms the baseline method in terms of voice quality and voice similarity.

Metaverse is an interactive world that combines reality and virtuality, where participants can be virtual avatars. Anyone can hold a concert in a virtual concert hall, and users can quickly identify the real singer behind the virtual idol through the singer identification. Most singer identification methods are processed using the frame-level features. However, expect the singer's timbre, the music frame includes music information, such as melodiousness, rhythm, and tonal. It means the music information is noise for using frame-level features to identify the singers. In this paper, instead of only the frame-level features, we propose to use another two features that address this problem. Middle-level feature, which represents the music's melodiousness, rhythmic stability, and tonal stability, and is able to capture the perceptual features of music. The timbre feature, which is used in speaker identification, represents the singers' voice features. Furthermore, we propose a convolutional recurrent neural network (CRNN) to combine three features for singer identification. The model firstly fuses the frame-level feature and timbre feature and then combines middle-level features to the mix features. In experiments, the proposed method achieves comparable performance on an average F1 score of 0.81 on the benchmark dataset of Artist20, which significantly improves related works.

Metaverse has stretched the real world into unlimited space. There will be more live concerts in Metaverse. The task of singer identification is to identify the song belongs to which singer. However, there has been a tough problem in singer identification, which is the different live effects. The studio version is different from the live version, the data distribution of the training set and the test set are different, and the performance of the classifier decreases. This paper proposes the use of the domain adaptation method to solve the live effect in singer identification. Three methods of domain adaptation combined with Convolutional Recurrent Neural Network (CRNN) are designed, which are Maximum Mean Discrepancy (MMD), gradient reversal (Revgrad), and Contrastive Adaptation Network (CAN). MMD is a distance-based method, which adds domain loss. Revgrad is based on the idea that learned features can represent different domain samples. CAN is based on class adaptation, it takes into account the correspondence between the categories of the source domain and target domain. Experimental results on the public dataset of Artist20 show that CRNN-MMD leads to an improvement over the baseline CRNN by 0.14. The CRNN-RevGrad outperforms the baseline by 0.21. The CRNN-CAN achieved state of the art with the F1 measure value of 0.83 on album split.

Speaker verification (SV) aims to determine whether the speaker's identity of a test utterance is the same as the reference speech. In the past few years, extracting speaker embeddings using deep neural networks for SV systems has gone mainstream. Recently, different attention mechanisms and Transformer networks have been explored widely in SV fields. However, utilizing the original Transformer in SV directly may have frame-level information waste on output features, which could lead to restrictions on capacity and discrimination of speaker embeddings. Therefore, we propose an approach to derive utterance-level speaker embeddings via a Transformer architecture that uses a novel loss function named diffluence loss to integrate the feature information of different Transformer layers. Therein, the diffluence loss aims to aggregate frame-level features into an utterance-level representation, and it could be integrated into the Transformer expediently. Besides, we also introduce a learnable mel-fbank energy feature extractor named time-domain feature extractor that computes the mel-fbank features more precisely and efficiently than the standard mel-fbank extractor. Combining Diffluence loss and Time-domain feature extractor, we propose a novel Transformer-based time-domain SV model (DT-SV) with faster training speed and higher accuracy. Experiments indicate that our proposed model can achieve better performance in comparison with other models.

Sound event detection is to infer the event by understanding the surrounding environmental sounds. Due to the scarcity of rare sound events, it becomes challenging for the well-trained detectors which have learned too much prior knowledge. Meanwhile, few-shot learning methods promise a good generalization abaility when facing a new limited-data task. Recent approaches have achieved promising results in this field. However, these approaches treat each support example independently, ignoring the information of other examples from the whole task. Because of this, most of previous methods are constrained to generate a same feature embedding for all test-time tasks, which is not adaptive to each inputted data. In this work, we propose a novel task-adaptive module which is easy to plant into any metric-based few-shot learning frameworks. The module could identify the task-relevant feature dimension. Incorporating our module improves the performance considerably on two datasets over baseline methods, especially for the transductive propagation network. Such as +6.8% for 5-way 1-shot accuracy on ESC-50, and +5.9% on noiseESC-50. We investigate our approach in the domain-mismatch setting and also achieve better results than previous methods.

The increasingly stringent regulations on privacy protection have sparked interest in federated learning. As a distributed machine learning framework, it bridges isolated data islands by training a global model over devices while keeping data localized. Specific to recommendation systems, many federated recommendation algorithms have been proposed to realize the privacy-preserving collaborative recommendation. However, several constraints remain largely unexplored. One big concern is how to ensure fairness between participants of federated learning, that is, to maintain the uniformity of recommendation performance across devices. On the other hand, due to data heterogeneity and limited networks, additional challenges occur in the convergence speed. To address these problems, in this paper, we first propose a personalized federated recommendation system training algorithm to improve the recommendation performance fairness. Then we adopt a clustering-based aggregation method to accelerate the training process. Combining the two components, we proposed Cali3F, a calibrated fast and fair federated recommendation framework. Cali3F not only addresses the convergence problem by a within-cluster parameter sharing approach but also significantly boosts fairness by calibrating local models with the global model. We demonstrate the performance of Cali3F across standard benchmark datasets and explore the efficacy in comparison to traditional aggregation approaches.

Speech emotion recognition (SER) has many challenges, but one of the main challenges is that each framework does not have a unified standard.In this paper, we propose SpeechEQ, a framework for unifying SER tasks based on a multi-scale unified metric.This metric can be trained by Multitask Learning (MTL), which includes two emotion recognition tasks of Emotion States Category (EIS) and Emotion Intensity Scale (EIS), and two auxiliary tasks of phoneme recognition and gender recognition.For this framework, we build a Mandarin SER dataset -SpeechEQ Dataset (SEQD).We conducted experiments on the public CASIA and ESD datasets in Mandarin, which exhibit that our method outperforms baseline methods by a relatively large margin, yielding 8.0% and 6.5% improvement in accuracy respectively.Additional experiments on IEMOCAP with four emotion categories (i.e., angry, happy, sad, and neutral) also show the proposed method achieves a state-of-the-art of both weighted accuracy (WA) of 78.16% and unweighted accuracy (UA) of 77.47%.

Time-domain Transformer neural networks have proven their superiority in speech separation tasks.However, these models usually have a large number of network parameters, thus often encountering the problem of GPU memory explosion.In this paper, we proposed Tiny-Sepformer, a tiny version of Transformer network for speech separation.We present two techniques to reduce the model parameters and memory consumption: (1) Convolution-Attention (CA) block, spliting the vanilla Transformer to two paths, multi-head attention and 1D depthwise separable convolution, (2) parameter sharing, sharing the layer parameters within the CA block.In our experiments, Tiny-Sepformer could greatly reduce the model size, and achieves comparable separation performance with vanilla Sepformer on WSJ0-2/3Mix datasets.

Although deep Neural Networks (DNNs) have achieved tremendous success in audio classification tasks, their uncertainty calibration are still under-explored.A well-calibrated model should be accurate when it is certain about its prediction and indicate high uncertainty when it is likely to be inaccurate.In this work, we investigate the uncertainty calibration for deep audio classifiers.In particular, we empirically study the performance of popular calibration methods: (i) Monte Carlo Dropout, (ii) ensemble, (iii) focal loss, and (iv) spectralnormalized Gaussian process (SNGP), on audio classification datasets.To this end, we evaluate (i-iv) for the tasks of environment sound and music genre classification.Results indicate that uncalibrated deep audio classifiers may be over-confident, and SNGP performs the best and is very efficient on the two datasets of this paper.

Multi-view stereopsis (MVS) recovers 3D surfaces by finding dense photo-consistent correspondences from densely sampled images. In this paper, we tackle the challenging MVS task from sparsely sampled views (up to an order of magnitude fewer images), which is more practical and cost-efficient in applications. The major challenge comes from the significant correspondence ambiguity introduced by the severe occlusions and the highly skewed patches. On the other hand, such ambiguity can be resolved by incorporating geometric cues from the global structure. In light of this, we propose ParseMVS, boosting sparse MVS by learning the P rimitive-A waR e S urface rE presentation. In particular, on top of being aware of global structure, our novel representation further allows for the preservation of fine details including geometry, texture, and visibility. More specifically, the whole scene is parsed into multiple geometric primitives. On each of them, the geometry is defined as the displacement along the primitives' normal directions, together with the texture and visibility along each view direction. An unsupervised neural network is trained to learn these factors by progressively increasing the photo-consistency and render-consistency among all input images. Since the surface properties are changed locally in the 2D space of each primitive, ParseMVS can preserve global primitive structures while optimizing local details, handling the 'incompleteness' and the 'inaccuracy' problems. We experimentally demonstrate that ParseMVS constantly outperforms the state-of-the-art surface reconstruction method in both completeness and the overall score under varying sampling sparsity, especially under the extreme sparse-MVS settings. Beyond that, ParseMVS also shows great potential in compression, robustness, and efficiency.

In this paper, we proposed Adapitch, a multi-speaker TTS method that makes adaptation of the supervised module with untranscribed data. We design two self supervised modules to train the text encoder and mel decoder separately with untranscribed data to enhance the representation of text and mel. To better handle the prosody information in a synthesized voice, a supervised TTS module is designed conditioned on content disentangling of pitch, text, and speaker. The training phase was separated into two parts, pretrained and fixed the text encoder and mel decoder with unsupervised mode, then the supervised mode on the disentanglement of TTS. Experiment results show that the Adaptich achieved much better quality than baseline methods.

Metaverse expands the physical world to a new dimension, and the physical environment and Metaverse environment can be directly connected and entered. Voice is an indispensable communication medium in the real world and Metaverse. Fusion of the voice with environment effects is important for user immersion in Metaverse. In this paper, we proposed using the voice conversion based method for the conversion of target environment effect speech. The proposed method was named MetaSpeech, which introduces an environment effect module containing an effect extractor to extract the environment information and an effect encoder to encode the environment effect condition, in which gradient reversal layer was used for adversarial training to keep the speech content and speaker information while disentangling the environmental effects. From the experiment results on the public dataset of LJSpeech with four environment effects, the proposed model could complete the specific environment effect conversion and outperforms the baseline methods from the voice conversion task.

Most previous neural text-to-speech (TTS) methods are mainly based on supervised learning methods, which means they depend on a large training dataset and hard to achieve comparable performance under low-resource conditions. To ad-dress this issue, we propose a semi-supervised learning method for neural TTS in which labeled target data is limited, which can also resolve the problem of exposure bias in the previous auto-regressive models. Specifically, we pre-train the reference model based on Fastspeech2 with much source data, fine-tuned on a limited target dataset. Meanwhile, pseudo labels generated by the original reference model are used to guide the fine-tuned model's training further, achieve a regularization effect, and reduce the overfitting of the fine-tuned model during training on the limited target data. Experimental results show that our proposed semi-supervised learning scheme with limited target data significantly improves the voice quality for test data to achieve naturalness and robustness in speech synthesis.

Recent advances in pre-trained language models have improved the performance for text classification tasks. However, little attention is paid to the priority scheduling strategy on the samples during training. Humans acquire knowledge gradually from easy to complex concepts, and the difficulty of the same material can also vary significantly in different learning stages. Inspired by this insights, we proposed a novel self-paced dynamic curriculum learning (SPDCL) method for imbalanced text classification, which evaluates the sample difficulty by both linguistic character and model capacity. Meanwhile, rather than using static curriculum learning as in the existing research, our SPDCL can reorder and resample training data by difficulty criterion with an adaptive from easy to hard pace. The extensive experiments on several classification tasks show the effectiveness of SPDCL strategy, especially for the imbalanced dataset.

Recovering the masked speech frames is widely applied in speech representation learning. However, most of these models use random masking in the pre-training. In this work, we proposed two kinds of masking approaches: (1) speech-level masking, making the model to mask more speech segments than silence segments, (2) phoneme-level masking, forcing the model to mask the whole frames of the phoneme, instead of phoneme pieces. We pre-trained the model via these two approaches, and evaluated on two downstream tasks, phoneme classification and speaker recognition. The experiments demonstrated that the proposed masking approaches are beneficial to improve the performance of speech representation.

The recent emergence of joint CTC-Attention model shows significant improvement in automatic speech recognition (ASR). The improvement largely lies in the modeling of linguistic information by decoder. The decoder joint-optimized with an acoustic encoder renders the language model from ground-truth sequences in an auto-regressive manner during training. However, the training corpus of the decoder is limited to the speech transcriptions, which is far less than the corpus needed to train an acceptable language model. This leads to poor robustness of decoder. To alleviate this problem, we propose linguistic-enhanced transformer, which introduces refined CTC information to decoder during training process, so that the decoder can be more robust. Our experiments on AISHELL-1 speech corpus show that the character error rate (CER) is relatively reduced by up to 7 %. We also find that in joint CTC-Attention ASR model, decoder is more sensitive to linguistic information than acoustic information.

Self-supervised multi-view stereopsis (MVS) attracts increasing attention for learning dense surface predictions from only a set of images without onerous ground-truth 3D training data for supervision. However, existing methods highly rely on the local photometric consistency, which fails to identify accurately dense correspondence in broad textureless and reflectance areas.In this paper, we show that geometric proximity such as surface connectedness and occlusion boundaries implicitly inferred from images could serve as reliable guidance for pixel-wise multi-view correspondences. With this insight, we present a novel elastic part representation which encodes physically-connected part segmentations with elastically-varying scales, shapes and boundaries. Meanwhile, a self-supervised MVS framework namely ElasticMVS is proposed to learn the representation and estimate per-view depth following a part-aware propagation and evaluation scheme. Specifically, the pixel-wise part representation is trained by a contrastive learning-based strategy, which increases the representation compactness in geometrically concentrated areas and contrasts otherwise. ElasticMVS iteratively optimizes a part-level consistency loss and a surface smoothness loss, based on a set of depth hypotheses propagated from the geometrically concentrated parts. Extensive evaluations convey the superiority of ElasticMVS in the reconstruction completeness and accuracy, as well as the efficiency and scalability. Particularly, for the challenging large-scale reconstruction benchmark, ElasticMVS demonstrates significant performance gain over both the supervised and self-supervised approaches.

Estimating age from a single speech is a classic and challenging topic. Although Label Distribution Learning (LDL) can represent adjacent indistinguishable ages well, the uncertainty of the age estimate for each utterance varies from person to person, i.e., the variance of the age distribution is different. To address this issue, we propose selective variance label distribution learning (SVLDL) method to adapt the variance of different age distributions. Furthermore, the model uses WavLM as the speech feature extractor and adds the auxiliary task of gender recognition to further improve the performance. Two tricks are applied on the loss function to enhance the robustness of the age estimation and improve the quality of the fitted age distribution. Extensive experiments show that the model achieves state-of-the-art performance on all aspects of the NIST SRE08-10 and a real-world datasets.

Unsupervised representation learning for speech audios attained impressive performances for speech recognition tasks, particularly when annotated speech is limited. However, the unsupervised paradigm needs to be carefully designed and little is known about what properties these representations acquire. There is no guarantee that the model learns meaningful representations for valuable information for recognition. Moreover, the adaptation ability of the learned representations to other domains still needs to be estimated. In this work, we explore learning domain-invariant representations via a direct mapping of speech representations to their corresponding high-level linguistic informations. Results prove that the learned latents not only capture the articulatory feature of each phoneme but also enhance the adaptation ability, outperforming the baseline largely on accented benchmarks.

Radiological images such as computed tomography (CT) and X-rays render anatomy with intrinsic structures. Being able to reliably locate the same anatomical structure across varying images is a fundamental task in medical image analysis. In principle it is possible to use landmark detection or semantic segmentation for this task, but to work well these require large numbers of labeled data for each anatomical structure and sub-structure of interest. A more universal approach would learn the intrinsic structure from unlabeled images. We introduce such an approach, called Self-supervised Anatomical eMbedding (SAM). SAM generates semantic embeddings for each image pixel that describes its anatomical location or body part. To produce such embeddings, we propose a pixel-level contrastive learning framework. A coarse-to-fine strategy ensures both global and local anatomical information are encoded. Negative sample selection strategies are designed to enhance the embedding's discriminability. Using SAM, one can label any point of interest on a template image and then locate the same body part in other images by simple nearest neighbor searching. We demonstrate the effectiveness of SAM in multiple tasks with 2D and 3D image modalities. On a chest CT dataset with 19 landmarks, SAM outperforms widely-used registration algorithms while only taking 0.23 seconds for inference. On two X-ray datasets, SAM, with only one labeled template image, surpasses supervised methods trained on 50 labeled images. We also apply SAM on whole-body follow-up lesion matching in CT and obtain an accuracy of 91%. SAM can also be applied for improving image registration and initializing CNN weights.

Object detection methods are widely adopted for computer-aided diagnosis using medical images. Anomalous findings are usually treated as objects that are described by bounding boxes. Yet, many pathological findings, e.g., bone fractures, cannot be clearly defined by bounding boxes, owing to considerable instance, shape and boundary ambiguities. This makes bounding box annotations, and their associated losses, highly ill-suited. In this work, we propose a new bone fracture detection method for X-ray images, based on a labor effective and flexible annotation scheme suitable for abnormal findings with no clear object-level spatial extents or boundaries. Our method employs a simple, intuitive, and informative point-based annotation protocol to mark localized pathology information. To address the uncertainty in the fracture scales annotated via point(s), we convert the annotations into pixel-wise supervision that uses lower and upper bounds with positive, negative, and uncertain regions. A novel Window Loss is subsequently proposed to only penalize the predictions outside of the uncertain regions. Our method has been extensively evaluated on 4410 pelvic X-ray images of unique patients. Experiments demonstrate that our method outperforms previous state-of-the-art image classification and object detection baselines by healthy margins, with an AUROC of 0.983 and FROC score of 89.6%.

Li Huang, Junjie Li, Weiwei Jiang, Zhiyu Zhang, Minchuan Chen, Shaojun Wang, Jing Xiao. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Zongcheng Ji, Tian Xia, Mei Han, Jing Xiao. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

We provide a theoretical analysis on how gradient clipping affects the convergence of the incremental gradient methods on minimizing an objective function that is the sum of a large number of component functions. We show that clipping on gradients of component functions leads to bias on the descent direction, which is affected by the clipping threshold, the norms of gradients of component functions, together with the angles between gradients of component functions and the full gradient. We then propose some sufficient conditions under which the increment gradient methods with gradient clipping can be shown to be convergent under the more general relaxed smoothness assumption. We also empirically observe that the angles between gradients of component functions and the full gradient generally decrease as the batchsize increases, which may help to explain why larger batchsizes generally lead to faster convergence in training deep neural networks with gradient clipping.

This paper introduces a dual learning system for neural voice conversion (DualVC) using relatively few samples based on the symmetry of the speech conversion task. The system contains a pair of sequence-to-sequence neural networks that have the same structure but are trained in opposite directions. The objective function of the dual model training is the sum of paired conversion loss and reconstruction loss during the dual training circle. The models in the two directions are trained alternately to guide each other by the corresponding reconstruction loss. Furthermore, curriculum learning techniques are used to load models in existing fields into the current task to accelerate the rapid iteration and convergence of the model. The experiment on the voice conversion task with the proposed DualVC and curriculum learning strategy obtained a comparable naturalness and similarity with only a 30% dataset than the BaseVC model trained on the full dataset.

Non-parallel many-to-many voice conversion remains an interesting but challenging speech processing task. Recently, AutoVC, a conditional autoencoder based method, achieved excellent conversion results by disentangling the speaker identity and the speech content using information-constraining bottlenecks. However, due to the pure autoencoder training method, it is difficult to evaluate the separation effect of content and speaker identity. In this paper, a novel voice conversion framework, named Text Guided AutoVC(TGAVC), is proposed to more effectively separate content and timbre from speech, where an expected content embedding produced based on the text transcriptions is designed to guide the extraction of voice content. In addition, the adversarial training is applied to eliminate the speaker identity information in the estimated content embedding extracted from speech. Under the guidance of the expected content embedding and the adversarial training, the content encoder is trained to extract speaker-independent content embedding from speech. Experiments on AIShell-3 dataset show that the proposed model outperforms AutoVC in terms of naturalness and similarity of converted speech.

Cycle Generative Adversarial Network (CycleGAN) for voice conversion (VC) task only used discriminators to identify whether the input voice is generated or real. It means the confrontational does not check the similarity with the target voice, leading the generated voice not much similar to the target. In this paper, instead of vocal checking, we propose to enhance the confrontation to target similarity checking that addresses this problem. A Cycle Generative Enhanced Adversarial Network (CycleGEAN) was introduced to make the original two discriminators to target classifier and non-target classifier. The target classifier aims to identify whether the target speaks the input voice or not. Similarly, the non-target classifier identifies the non-target voice. Furthermore, we add a gradient reversal layer with different operations for target and non-target. Then in each GAN, we used both classifiers. One is the discriminator, and the other is trained for using in another GAN. In experiments, the proposed method compare to CycleGAN improves Mean Opinion Score (MOS) of 0.1 and Voice Similarity Score (VSS) of 0.2 on the Voice Conversion Challenge 2018 (VCC2018) dataset.

Monitoring treatment response in longitudinal studies plays an important role in clinical practice. Accurately identifying lesions across serial imaging follow-up is the core to the monitoring procedure. Typically this incorporates both image and anatomical considerations. However, matching lesions manually is labor-intensive and time-consuming. In this work, we present deep lesion tracker (DLT), a deep learning approach that uses both appearance- and anatomical-based signals. To incorporate anatomical constraints, we propose an anatomical signal encoder, which prevents lesions being matched with visually similar but spurious regions. In addition, we present a new formulation for Siamese networks that avoids the heavy computational loads of 3D cross-correlation. To present our network with greater varieties of images, we also propose a self-supervised learning (SSL) strategy to train trackers with unpaired images, overcoming barriers to data collection. To train and evaluate our tracker, we introduce and release the first lesion tracking benchmark, consisting of 3891 lesion pairs from the public DeepLesion database. The proposed method, DLT, locates lesion centers with a mean error distance of 7mm. This is 5% better than a leading registration algorithm while running 14 times faster on whole CT volumes. We demonstrate even greater improvements over detector or similarity-learning alternatives. DLT also generalizes well on an external clinical test set of 100 longitudinal studies, achieving 88% accuracy. Finally, we plug DLT into an automatic tumor monitoring workflow where it leads to an accuracy of 85% in assessing lesion treatment responses, which is only 0.46% lower than the accuracy of manual inputs.

Mammographic mass detection is an integral part of a computer-aided diagnosis system. Annotating a large number of mammograms at pixel-level in order to train a mass detection model in a fully supervised fashion is costly and time-consuming. This paper presents a novel self-training framework for semi-supervised mass detection with soft image-level labels generated from diagnosis reports by Mammo-RoBERTa, a RoBERTa-based natural language processing model fine-tuned on the fully labeled data and associated mammography reports. Starting with a fully supervised model trained on the data with pixel-level masks, the proposed framework iteratively refines the model itself using the entire weakly labeled data (image-level soft label) in a self-training fashion. A novel sample selection strategy is proposed to identify those most informative samples for each iteration, based on the current model output and the soft labels of the weakly labeled data. A soft cross-entropy loss and a soft focal loss are also designed to serve as the image-level and pixel-level classification loss respectively. Our experiment results show that the proposed semi-supervised framework can improve the mass detection accuracy on top of the supervised baseline, and outperforms the previous state-of-the-art semi-supervised approaches with weakly labeled data, in some cases by a large margin.

Accurate vertebra localization and identification are required in many clinical applications of spine disorder diagnosis and surgery planning. However, significant challenges are posed in this task by highly varying pathologies (such as vertebral compression fracture, scoliosis, and vertebral fixation) and imaging conditions (such as limited field of view and metal streak artifacts). This paper proposes a robust and accurate method that effectively exploits the anatomical knowledge of the spine to facilitate vertebra localization and identification. A key point localization model is trained to produce activation maps of vertebra centers. They are then re-sampled along the spine centerline to produce spine-rectified activation maps, which are further aggregated into 1-D activation signals. Following this, an anatomically-constrained optimization module is introduced to jointly search for the optimal vertebra centers under a soft constraint that regulates the distance between vertebrae and a hard constraint on the consecutive vertebra indices. When being evaluated on a major public benchmark of 302 highly pathological CT images, the proposed method reports the state of the art identification (id.) rate of 97.4%, and outperforms the best competing method of 94.7% id. rate by reducing the relative id. error rate by half.

The pancreatic disease taxonomy includes ten types of masses (tumors or cysts) [20], [8]. Previous work focuses on developing segmentation or classification methods only for certain mass types. Differential diagnosis of all mass types is clinically highly desirable [20] but has not been investigated using an automated image understanding approach.We exploit the feasibility to distinguish pancreatic ductal adenocarcinoma (PDAC) from the nine other nonPDAC masses using multi-phase CT imaging. Both image appearance and the 3D organ-mass geometry relationship are critical. We propose a holistic segmentation-mesh-classification network (SMCN) to provide patient-level diagnosis, by fully utilizing the geometry and location information, which is accomplished by combining the anatomical structure and the semantic detection-by-segmentation network. SMCN learns the pancreas and mass segmentation task and builds an anatomical correspondence-aware organ mesh model by progressively deforming a pancreas prototype on the raw segmentation mask (i.e., mask-to-mesh). A new graph-based residual convolutional network (Graph-ResNet), whose nodes fuse the information of the mesh model and feature vectors extracted from the segmentation network, is developed to produce the patient-level differential classification results. Extensive experiments on 661 patients’ CT scans (five phases per patient) show that SMCN can improve the mass segmentation and detection accuracy compared to the strong baseline method nnUNet (e.g., for nonPDAC, Dice: 0.611 vs. 0.478; detection rate: 89% vs. 70%), achieve similar sensitivity and specificity in differentiating PDAC and nonPDAC as expert radiologists (i.e., 94% and 90%), and obtain results comparable to a multimodality test [20] that combines clinical, imaging, and molecular testing for clinical management of patients.

Inspired by the contrastive predictive coding (CPC), we propose a feature representation scheme for automatic speech recognition (ASR), which encodes sequential dependency information from raw audio signals. Following the original CPC, for a given frame, mutual information (MI) lower bound is maximized between historical context and future prediction. While computing the MI lower bound, based on original CPC, we develop the sequential CPC (SEQ-CPC), which takes the sequential information between frames into consideration. Since speech frames are not independent events, incorporating sequential information leads to better recognition performance. Experimental results on WSJ corpus show that SEQ-CPC achieves the best performance than CPC and NCE which is the contrastive objective used in wav2vec.

We propose a CTC alignment-based single step non-autoregressive transformer (CASS-NAT) for speech recognition. Specifically, the CTC alignment contains the information of (a) the number of tokens for decoder input, and (b) the time span of acoustics for each token. The information are used to extract acoustic representation for each token in parallel, referred to as token-level acoustic embedding which substitutes the word embedding in autoregressive transformer (AT) to achieve parallel generation in decoder. During inference, an error-based alignment sampling method is proposed to be applied to the CTC output space, reducing the WER and retaining the parallelism as well. Experimental results show that the proposed method achieves WERs of 3.8%/9.1% on Librispeech test clean/other dataset without an external LM, and a CER of 5.8% on Aishell1 Mandarin corpus, respectively1. Compared to the AT baseline, the CASS-NAT has a performance reduction on WER, but is 51.2x faster in terms of RTF. When decoding with an oracle CTC alignment, the lower bound of WER without LM reaches 2.3% on the test-clean set, indicating the potential of the proposed method.

Slot filling and intent detection have become a significant theme in the field of natural language understanding. Even though slot filling is intensively associated with intent detection, the characteristics of the information required for both tasks are different while most of those approaches may not fully aware of this problem. In addition, balancing the accuracy of two tasks effectively is an inevitable problem for the joint learning model. In this paper, a Continual Learning Interrelated Model (CLIM) is proposed to consider semantic information with different characteristics and balance the accuracy between intent detection and slot filling effectively. The experimental results show that CLIM achieves state-of-the-art performace on slot filling and intent detection on ATIS and Snips.

Text Normalization (TN) is an essential part in conversational systems like text-to-speech synthesis (TTS) and automatic speech recognition (ASR). It is a process of transforming non-standard words (NSW) into a representation of how the words are to be spoken. Existing approaches to TN are mainly rule-based or hybrid systems, which require abundant hand-crafted rules. In this paper, we treat TN as a neural machine translation problem and present a pure data-driven TN system using Transformer framework. Partial Parameter Generator (PPG) and Pointer-Generator Network (PGN) are combined in our model to improve accuracy of normalization and act as auxiliary modules to reduce the number of simple errors. The experiments demonstrate that our proposed model reaches remarkable performance on various semiotic classes.

Existing multi-style speech synthesis methods require either style labels or large amounts of unlabeled training data, making data acquisition difficult. In this paper, we present an unsupervised multi-style speech synthesis method that can be trained with limited data. We leverage instance discriminator to guide a style encoder to learn meaningful style representations from a multi-style dataset. Furthermore, we employ information bottleneck to filter out style-irrelevant information in the representations, which can improve speech quality and style similarity. Our method is able to produce desirable speech using a fairly small dataset, where the baseline GST-Tacotron fails. ABX tests show that our model significantly outperforms GST-Tacotron in both emotional speech synthesis task and multi-speaker speech synthesis task. In addition, we demonstrate that our method is able to learn meaningful style features with only 50 training samples per style.

In the federated learning system, parameter gradients are shared among participants and the central modulator, while the original data never leave their protected source domain. However, the gradient itself might carry enough information for precise inference of the original data. By reporting their parameter gradients to the central server, client datasets are exposed to inference attacks from adversaries. In this paper, we propose a quantitative metric based on mutual information for clients to evaluate the potential risk of information leakage in their gradients. Mutual information has received increasing attention in the machine learning and data mining community over the past few years. However, existing mutual information estimation methods cannot handle high-dimensional variables. In this paper, we propose a novel method to approximate the mutual information between the high-dimensional gradients and batched input data. Experimental results show that the proposed metric reliably reflect the extent of information leakage in federated learning. In addition, using the proposed metric, we investigate the influential factors of risk level. It is proven that, the risk of information leakage is related to the status of the task model, as well as the inherent data distribution.

Self-attention models have been successfully applied in end-to-end speech recognition systems, which greatly improve the performance of recognition accuracy. However, such attention-based models cannot be used in online speech recognition, because these models usually have to utilize a whole acoustic sequences as inputs. A common method is restricting the field of attention sights by a fixed left and right window, which makes the computation costs manageable yet also introduces performance degradation. In this paper, we propose Memory-Self-Attention (MSA), which adds history information into the Restricted-Self-Attention unit. MSA only needs localtime features as inputs, and efficiently models long temporal contexts by attending memory states. Meanwhile, recurrent neural network transducer (RNN-T) has proved to be a great approach for online ASR tasks, because the alignments of RNN-T are local and monotonic. We propose a novel network structure, called Memory-Self-Attention (MSA) Transducer. Both encoder and decoder of the MSA Transducer contain the proposed MSA unit. The experiments demonstrate that our proposed models improve WER results than Restricted-Self-Attention models by 13.5% on WSJ and 7.1% on SWBD datasets relatively, and without much computation costs increase.

Network pruning can be achieved by removing redundant channels. In this paper, we regard a channel ‘redundant’ if its output is linearly dependent with respect to those of other channels. Inspired by this, we propose an efficient pruning method, named as LDFM, by linear dependency analysis on all the feature maps of each individual layer. Specifically, for each layer, by applying the QR decomposition with column pivoting (PQR) on the matrix consisting of all feature maps, those channels corresponding to small absolute diagonal elements of the R matrix from the PQR decomposition are identified as redundant, and are pruned naturally. Although pruning these channels causes loss of information and hence degrades accuracy, the accuracy of the pruned network can be easily recovered by fine-tuning, as the lost information in the pruned channels can be recovered from that in the retained channels. Extensive experiments demonstrate that LDFM makes great improvement on accuracy with similar parameters and FLOPs as other methods, and achieves the state-of-the-art results on several different benchmarks and networks.

Knowledge distillation refers to a technique of transferring the knowledge from a large learned model or an ensemble of learned models to a small model. This method relies on access to the original training set, which might not always be available. A possible solution is a data-free adversarial distillation framework, which deploys a generative network to transfer the teacher model’s knowledge to the student model. However, the data generation efficiency is low in the data-free adversarial distillation. We add an activation regularizer and a virtual interpolation method to improve the data generation efficiency. The activation regularizer enables the students to match the teacher’s predictions close to activation boundaries and decision boundaries. The virtual interpolation method can generate virtual samples and labels in-between decision boundaries. Our experiments show that our approach surpasses state-of-the-art data-free distillation methods. The student model can achieve 95.42% accuracy on CIFAR-10 and 77.05% accuracy on CIFAR-100 without any original training data. Our model’s accuracy is 13.8% higher than the state-of-the-art data-free method on CIFAR-100.

In this paper, we propose a novel conditional convolution network, named location-variable convolution, to model the dependencies of the waveform sequence. Different from the use of unified convolution kernels in WaveNet to capture the dependencies of arbitrary waveform, the location-variable convolution uses convolution kernels with different coefficients to perform convolution operations on different waveform intervals, where the coefficients of kernels is predicted according to conditioning acoustic features, such as Mel-spectrograms. Based on location-variable convolutions, we design LVCNet for waveform generation, and apply it in Parallel WaveGAN to design more efficient vocoder. Experiments on the LJSpeech dataset show that our proposed model achieves a four-fold increase in synthesis speed compared to the original Parallel WaveGAN without any degradation in sound quality, which verifies the effectiveness of location-variable convolutions.

In federated learning (FL), fair and accurate measurement of the contribution of each federated participant is of great significance. The level of contribution not only provides a rational metric for distributing financial benefits among federated participants, but also helps to discover malicious participants that try to poison the FL framework. Previous methods for contribution measurement were based on enumeration over possible combination of federated participants. Their computation costs increase drastically with the number of participants or feature dimensions, making them inapplicable in practical situations. In this paper an efficient method is proposed to evaluate the contributions of federated participants. This paper focuses on the horizontal FL framework, where client servers calculate parameter gradients over their local data, and upload the gradients to the central server. Before aggregating the client gradients, the central server train a data value estimator of the gradients using reinforcement learning techniques. As shown by experimental results, the proposed method consistently outperforms the conventional leave-one-out method in terms of valuation authenticity as well as time complexity.

Network embedding, which aims to map the discrete network topology to a continuous low-dimensional representation space with the major topological properties preserved, has emerged as an essential technique to support various network inference tasks. However, incorporating both the evolutionary nature and the network's heterogeneity remains a challenge for existing network embedding methods. In this study, we propose a novel Translation-Based Dynamic Heterogeneous Network Embedding (TransDHE) approach to consider both the aspects simultaneously. For a dynamic heterogeneous network with a sequence of snapshots and multiple types of nodes and edges, we introduce a translation-based embedding module to capture the heterogeneous characteristics (e.g., type information) of each single snapshot. An orthogonal alignment module and RNN-based aggregation module are then applied to explore the evolutionary patterns among multiple successive snapshots for the final representation learning. Extensive experiments on a set of real-world networks demonstrate that TransDHE can derive the more informative embedding result for the network dynamic and heterogeneity over state-of-the-art network embedding baselines.

In this paper, we demonstrate the efficacy of transfer learning and continuous learning for various automatic speech recognition (ASR) tasks using end-to-end models trained with CTC loss. We start with a large pre-trained English ASR model and show that transfer learning can be effectively and easily performed on: (1) different English accents, (2) different languages (from English to German, Spanish, Russian, or from Mandarin to Cantonese) and (3) application-specific domains. Our extensive set of experiments demonstrate that in all three cases, transfer learning from a good base model has higher accuracy than a model trained from scratch. Our results indicate that, for fine-tuning, larger pre-trained models are better than small pre-trained models, even if the dataset for fine-tuning is small. We also show that transfer learning significantly speeds up convergence, which could result in significant cost savings when training with large datasets.

In this work, we address the Text-to-Speech (TTS) task by proposing a non-autoregressive architecture called EfficientTTS. Unlike the dominant non-autoregressive TTS models, which are trained with the need of external aligners, EfficientTTS optimizes all its parameters with a stable, end-to-end training procedure, allowing for synthesizing high quality speech in a fast and efficient manner. EfficientTTS is motivated by a new monotonic alignment modeling approach, which specifies monotonic constraints to the sequence alignment with almost no increase of computation. By combining EfficientTTS with different feed-forward network structures, we develop a family of TTS models, including both text-to-melspectrogram and text-to-waveform networks. We experimentally show that the proposed models significantly outperform counterpart models such as Tacotron 2 and Glow-TTS in terms of speech quality, training efficiency and synthesis speed, while still producing the speeches of strong robustness and great diversity. In addition, we demonstrate that proposed approach can be easily extended to autoregressive models such as Tacotron 2.

Proteins are linear polymers that fold into an incredible variety of three-dimensional structures that enable sophisticated functionality for biology. Predicting protein distance with high precision remains challenging, particularly for small protein families. As deep learning achieves remarkable success in many areas, deep learning also allows scientists to predict proteins' three-dimensional structure. As convolutional neural networks have a powerful ability to learn data features at multiple levels of abstraction, we deploy CNN to predict protein distance. To accelerate the training process, we apply a quantum convolutional neural network(QCNN) to improve the protein structure prediction efficiently. For QCNN, the conventional convolutional layer is transformed to a quantum convolution or quanvolutional layer. Since the protein data has large input, we explore the large dimension of these quantum transformations. And in experiments, we compare the different number of layers in QCNN during the training phase. We found the QCNN is similar to CNN that is the deeper layer can get better performance. The simulations show the proposed method can accelerate the convergence while maintaining the performance.

Speech is a carrier containing rich biological information, such as speaker identity information including age, gender, race. In this paper, we explore the use of a self-supervised method to obtain speaker identity information from high-dimensional speech representations to generate face image. At the same time, considering that the biological information contained in the same piece of speech has different expression forms (such as images), we designed a cross-modal knowledge distillation method to transform the feature information from the visual domain to the speech domain. The feature vectors obtained through self-supervised learning and knowledge distillation are fed into a GAN-based generative model to obtain facial images containing speaker information. Subjective experiments show that our model can reach a well performance in the task of speaker identification. Experiments show that our proposed method can effectively establish the connection between different modalities and generate a face with rich biological information.

Smartphones and wearable devices produce a wealth of audio data, which cannot be accumulated in a centralized repository for learning supervised models due to privacy and bandwidth limitation. Federated learning provides a solution for learning model from decentralized data. But conventionally, it assumes the availability of labeled samples, whereas on-device data are generally unlabeled. For solving these issues, in this paper we propose the self-supervised learning approach in a federated manner without moving the unlabeled audio data. We try the audio albert as the self-supervised model, which achieves comparable performance to other pre-trained model but with smaller model size. The federated self-supervised framework has tremendous communication cost during training, and the transformer architecture utilized in audio albert has the problem of memory footprint, which are practical in loT devices. To address the first issue, we propose the Gradient Compression and CSR Encoding (GCE) to reduce communication requires each round. Furthermore, we apply the reversible idea to the transformer, which does not need to store the activation in each layer thus reduce the memory footprint. Moreover, we evaluate the quality of the self-supervised pre-training model under the federated setting, and the model achieves considerable performance in the downstream tasks by fine-tuning.

Federated learning is a distributed deep learning method that enables parallel and distributed learning of data on multiple participants, without the need to centrally store it. In intelligent transportation system, it is impractical to gather the vehicle data from the edge devices due to data privacy concerns or network bandwidth limitation. Hence, combining with federated learning to train vehicle data processing models has become one of the popular solutions. However, such computing paradigm is subject to threats posed by malicious and abnormal nodes that greatly reduces the computing power of the neural network when performing calculations in a distributed manner. In this paper, we use the Vehicle Energy Dataset to simulate distributed vehicle data. Based on these data, we propose an unsupervised anomaly removal and neural network model based on federated learning to solve the problem of outlier data on vehicle equipment and analyze the effect of speed on fuel consumption. The results show that with the proposed anomaly removal strategy, MAE and MSE of the trained network are 29% and 36% lower than those without anomaly removal, respectively.

More accurate machine learning models often require more memory cost and more software-hardware co-adaption efforts for deployments on resource-constrained devices. Model compression techniques and deep learning compiler are developed to reduce the memory cost and latency. However, current methods require tremendous engineering efforts to optimize the model manually. This paper introduces a jointly learning based framework to perform the compression task and the acceleration task simultaneously. The joint optimization method auto-tunes the algorithm-level compression and compiler-based acceleration with reinforcement learning. The experiment results demonstrate that we compress the model by a factor of 2 or 8, and accelerate the optimization up to 30 times using our learning framework.

End-to-end speech recognition systems usually require huge amounts of labeling resource, while annotating the speech data is complicated and expensive. Active learning is the solution by selecting the most valuable samples for annotation. In this paper, we proposed to use a predicted loss that estimates the uncertainty of the sample. The CTC (Connectionist Temporal Classification) and attention loss are informative for speech recognition since they are computed based on all decoding paths and alignments. We defined an end-to-end active learning pipeline, training an ASR/LP (Automatic Speech Recognition/Loss Prediction) joint model. The proposed approach was validated on an English and a Chinese speech recognition task. The experiments show that our approach achieves competitive results, outperforming random selection, least confidence, and estimated loss method.

Speaker verification involves examining the speech signal to authenticate the claim of a speaker as true or false. Deep neural networks are one of the successful implementations of complex non-linear models to learn unique and invariant features of data. They have been employed in speech recognition tasks and have shown their potential to be used for speaker recognition also. However, the overfitting problem is remained to prevent the model's performance. In this study, we apply contrastive learning on speaker verification tasks to solve the robustness problem. Besides, we introduce domain adaptive loss on the tasks. Experimental results and ablation study that indicate that our proposed model outperforms various baseline end-to-end methods significantly by at least relative 10%, including d-vector approaches, deep-speaker, and generalized end-to-end model, for text-dependent speaker verification on a company's internal text-dependent voice command DataSet.

As a comprehensive and unstructured data source describing 3 dimensional (3D) objects, point cloud data has attracted wide interest from researchers. Neural networks (NNs) provide flexible, end-to-end solutions to point cloud processing tasks. Yet, one of the unavoidable problems in training the NNs is the difficulty in collecting and labeling high-quality point cloud data. Existing point cloud datasets are often collected at different scales and densities. Directly combining highly heterogeneous datasets might violate data protection regulations, but also lead to degradation of model performance. In our experiments, naive horizontal federated learning has been applied to alleviate the data privacy concern. However, when the datasets are not identically distributed, the federated model's performance falls far behind the state-of-the-art of centralized models. This paper investigates the union of highly heterogeneous point cloud datasets for the object classification task. In the proposed framework, the federated model is split into shared and personalized modules. The former improves generalizability by bringing multiple data sources into play, while the latter adapts local distribution and avoids convergence at false optima. Experiments show that the proposed framework FedPCN achieves 35.1% higher accuracy than FedAvg. At local test samples, FedPCN also outperforms the locally trained model by 0.5%.

Sequential recommendation system's goal is to predict users' next actions based on their historical behavior sequences. As we know, more recent items have a larger impact than the previous ones. Meanwhile, modeling users' current interests are challenging. Knowledge graph (KG) contains a vast of information, which can help us capture users' interests by propagating their interactions. In this paper, we propose an Attentive Dynamic Knowledge Graph Networks (ADKGN) for sequential recommendation, which includes an embedding module, an attentive dynamic knowledge graph module, a sequential process module and a prediction module. Specifically, embedding module learns initial item embedding vectors by combing latent features and sequential features. For users' recent interacted items, attentive dynamic knowledge graph module learns dynamic weights for each pretrained KG embedding, and then utilizes top-k layer and parallel-based aggregation layer to effectively aggregate useful information from multi-hop neighbors. Sequential process module combines a user's history interactions and processed current interactions, and employs sequential models over them to get final user representations. Prediction module predicts the clicking probability using final user representation and target item representation. We conduct experiments on a public dataset, finding that ADKGN significantly outperforms state-of-the-art solutions.

Many modern Natural Language Processing(NLP) systems rely on word embeddings, previously trained in an unsupervised manner on large corpora, as base features. Several attempts of learning unsupervised sentence representations have not achieved satisfactory performance and have not been widely adopted. In this work, we present a Semantic Extraction Reinforcement learning Model(SERM) for encoding sentences into embedding vectors, which transforms the learning process into intent detection and named entity recognition tasks. Intent detection is mainly related to the semantics of the whole sentence, while named entity recognition pays more attention to local entities. Given a sentence, SERM builds a sentence representation by extracting the most important words and removing irrelevant words in a sentence. Unlike the mainstream approach, SERM compresses sentences from semantic and entity perspectives and this allows us to efficiently learn different types of encoding functions. The experimental results show that SERM can learn high quality sentence representation. This paper demonstrates that our sentence representations have sufficient competitiveness with the best performing model on text classification task.

End-to-end models have been widely used in Automatic Speech Recognition (ASR). Convolutional Neural Networks (CNNs) can effectively use spectrum information to model acoustic models. However, the convolution layers have limitations on the receptive field leading to restrictions for long speech signals. Inspired by this, we propose a Cube Attention CNN network(CACnet) that uses two different attention blocks to integrate the feature information of different dimensions for extending context information. Thereinto, the Global Deep Attention Block utilizes non-local operations to compute interactions between any two positions on feature maps and enables the acquirement of global feature representations while the Cross-Channel Attention Block adaptively recalibrates channel-wise feature responses. Then, outputs of the above two attention modules will be added up to further improve the feature representation which contributes to enrich contextual information. Finally, the performance of our proposed architecture will be explored under ASR tasks in English circumstances. Experiments on LibriSpeech indicate that CACnet achieves a word error rate (WER) of 3.78%/9.56% without language model (LM), and 2.84%/6.97% with LM, which is near state-of-the-art accuracy. CACnet on WSJ with 4.4% WER obtains better performance, compared to CTC-based CNN models, such as QuartzNet and Jasper, with the same language model. The proposed network achieves competitive accuracy while having fewer parameters. Moreover, CACnet can be easily incorporated into any existed network since it has the same input and output dimensions.

Non-autoregressive mechanisms can significantly decrease inference time for speech transformers, especially when the single step variant is applied. Previous work on CTC alignment-based single step non-autoregressive transformer (CASS-NAT) has shown a large real time factor (RTF) improvement over autoregressive transformers (AT). In this work, we propose several methods to improve the accuracy of the end-to-end CASS-NAT, followed by performance analyses. First, convolution augmented self-attention blocks are applied to both the encoder and decoder modules. Second, we propose to expand the trigger mask (acoustic boundary) for each token to increase the robustness of CTC alignments. In addition, iterated loss functions are used to enhance the gradient update of low-layer parameters. Without using an external language model, the WERs of the improved CASS-NAT, when using the three methods, are 3.1%/7.2% on Librispeech test clean/other sets and the CER is 5.4% on the Aishell1 test set, achieving a 7%~21% relative WER/CER improvement. For the analyses, we plot attention weight distributions in the decoders to visualize the relationships between token-level acoustic embeddings. When the acoustic embeddings are visualized, we find that they have a similar behavior to word embeddings, which explains why the improved CASS-NAT performs similarly to AT.

Text to speech (TTS) is a crucial task for user interaction, but TTS model training relies on a sizable set of high-quality original datasets. Due to privacy and security issues, the original datasets are usually unavailable directly. Recently, federated learning proposes a popular distributed machine learning paradigm with an enhanced privacy protection mechanism. It offers a practical and secure framework for data owners to collaborate with others, thus obtaining a better global model trained on the larger dataset. However, due to the high complexity of transformer models, the convergence process becomes slow and unstable in the federated learning setting. Besides, the transformer model trained in federated learning is costly communication and limited computational speed on clients, impeding its popularity. To deal with these challenges, we propose the federated dynamic transformer. On the one hand, the performance is greatly improved comparing with the federated transformer, approaching centralize-trained Transformer-TTS when increasing clients number. On the other hand, it achieves faster and more stable convergence in the training phase and significantly reduces communication time. Experiments on the LJSpeech dataset also strongly prove our method's advantage.

Predicting the altered acoustic frames is an effective way of self-supervised learning for speech representation.However, it is challenging to prevent the pretrained model from overfitting.In this paper, we proposed to introduce two dropout regularization methods into the pretraining of transformer encoder: (1) attention dropout, (2) layer dropout.Both of the two dropout methods encourage the model to utilize global speech information, and avoid just copying local spectrum features when reconstructing the masked frames.We evaluated the proposed methods on phoneme classification and speaker recognition tasks.The experiments demonstrate that our dropout approaches achieve competitive results, and improve the performance of classification accuracy on downstream tasks.

This paper investigates a novel task of talking face video generation solely from speeches.The speech-to-video generation technique can spark interesting applications in entertainment, customer service, and human-computer-interaction industries.Indeed, the timbre, accent and speed in speeches could contain rich information relevant to speakers' appearance.The challenge mainly lies in disentangling the distinct visual attributes from audio signals.In this article, we propose a light-weight, cross-modal distillation method to extract disentangled emotional and identity information from unlabelled video inputs.The extracted features are then integrated by a generative adversarial network into talking face video clips.With carefully crafted discriminators, the proposed framework achieves realistic generation results.Experiments with observed individuals demonstrated that the proposed framework captures the emotional expressions solely from speeches, and produces spontaneous facial motion in the video output.Compared to the baseline method where speeches are combined with a static image of the speaker, the results of the proposed framework is almost indistinguishable.User studies also show that the proposed method outperforms the existing algorithms in terms of emotion expression in the generated videos.

Large-scale deep neural networks (DNNs) such as convolutional neural networks (CNNs) have achieved impressive performance in audio classification for their powerful capacity and strong generalization ability. However, when training a DNN model on low-resource tasks, it is usually prone to overfitting the small data and learning too much redundant information. To address this issue, we propose to use variational information bottleneck (VIB) to mitigate overfitting and suppress irrelevant information. In this work, we conduct experiments on a 4-layer CNN. However, the VIB framework is ready-to-use and could be easily utilized with many other state-of-the-art network architectures. Evaluation on a few audio datasets shows that our approach significantly outperforms baseline methods, yielding _ 5:0% improvement in terms of classification accuracy in some low-source settings. Copyright © 2021 ISCA.

Xuan Zhou, Xiao Zhang, Chenyang Tao, Junya Chen, Bing Xu, Wei Wang, Jing Xiao. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Different from traditional task-oriented and open-domain dialogue systems, insurance agents aim to engage customers for helping them satisfy specific demands and emotional companionship. As a result, customer-to-agent dialogues are usually very long, and many turns of them are pure chit-chat without any useful marketing clues. This brings challenges to dialogue state tracking task in insurance marketing. To deal with these long and sparse dialogues, we propose a new dialogue state tracking architecture containing three components: dialogue encoder, Smart History Collector (SHC) and dialogue state classifier. SHC, a deliberately designed memory network, effectively selects relevant dialogue history via slot-attention, and then updates dialogue history memory. With SHC, our model is able to keep track of the vital information and filter out pure chit-chat. Experimental results demonstrate that our proposed LS-DST significantly outperforms the state-of-the-art baselines on real insurance dialogue dataset.

In this paper, we propose an end-to-end speech recognition network based on Nvidia's previous QuartzNet [1] model. We try to promote the model performance, and design three components: (1) Multi-Resolution Convolution Module, re-places the original 1D time-channel separable convolution with multi-stream convolutions. Each stream has a unique dilated stride on convolutional operations. (2) Channel-Wise Attention Module, calculates the attention weight of each convolutional stream by spatial channel-wise pooling. (3) Multi-Layer Feature Fusion Module, reweights each convolutional block by global multi-layer feature maps. Our experiments demonstrate that Multi-QuartzNet model achieves CER 6.77% on AISHELL-1 data set, which outperforms original QuartzNet and is close to state-of-art result.

Silent speech interfaces (SSI) has been an exciting area of recent interest. In this paper, we present a non-invasive silent speech interface that uses inaudible acoustic signals to capture people's lip movements when they speak. We exploit the speaker and microphone of the smartphone to emit signals and listen to their reflections, respectively. The extracted phase features of these reflections are fed into the deep learning networks to recognize speech. And we also propose an end-to-end recognition framework, which combines the CNN and attention-based encoder-decoder network. Evaluation results on a limited vocabulary (54 sentences) yield word error rates of 8.4% in speaker-independent and environment-independent settings, and 8.1% for unseen sentence testing.

This paper introduces a graphical representation approach of prosody boundary (GraphPB) in the task of Chinese speech synthesis, intending to parse the semantic and syntactic relationship of input sequences in a graphical domain for improving the prosody performance. The nodes of the graph embedding are formed by prosodic words, and the edges are formed by the other prosodic boundaries, namely prosodic phrase boundary (PPH) and intonation phrase boundary (IPH). Different Graph Neural Networks (GNN) like Gated Graph Neural Network (GGNN) and Graph Long Short-term Memory (G-LSTM) are utilised as graph encoders to exploit the graphical prosody boundary information. Graph-to-sequence model is proposed and formed by a graph encoder and an attentional decoder. Two techniques are proposed to embed sequential information into the graph-to-sequence text-to-speech model. The experimental results show that this proposed approach can encode the phonetic and prosody rhythm of an utterance. The mean opinion score (MOS) of these GNN models shows comparative results with the state-of-the-art sequence-to-sequence models with better performance in the aspect of prosody. This provides an alternative approach for prosody modelling in end-to-end speech synthesis.

Recent neural vocoders usually use a WaveNet-like network to capture the long-term dependencies of the waveform, but a large number of parameters are required to obtain good modeling capabilities. In this paper, an efficient network, named location-variable convolution, is proposed to model the dependencies of waveforms. Different from the use of unified convolution kernels in WaveNet to capture the dependencies of arbitrary waveforms, location-variable convolutions utilizes a kernel predictor to generate multiple sets of convolution kernels based on the melspectrum, where each set of convolution kernels is used to perform convolution operations on the associated waveform intervals. Combining WaveGlow and location-variable convolutions, an efficient vocoder, named MelGlow, is designed. Experiments on the LJSpeech dataset show that MelGlow achieves better performance than WaveGlow at small model sizes, which verifies the effectiveness and potential optimization space of location-variable convolutions.

Many existing approaches for mammogram analysis are based on single view. Some recent DNN-based multi-view approaches can perform either bilateral or ipsilateral analysis, while in practice, radiologists use both to achieve the best clinical outcome. MommiNet is the first DNN-based tri-view mass identification approach, which can simultaneously perform bilateral and ipsilateral analysis of mammographic images, and in turn, can fully emulate the radiologists' reading practice. In this paper, we present MommiNet-v2, with improved network architecture and performance. Novel high-resolution network (HRNet)-based architectures are proposed to learn the symmetry and geometry constraints, to fully aggregate the information from all views for accurate mass detection. A multi-task learning scheme is adopted to incorporate both Breast Imaging-Reporting and Data System (BI-RADS) and biopsy information to train a mass malignancy classification network. Extensive experiments have been conducted on the public DDSM (Digital Database for Screening Mammography) dataset and our in-house dataset, and state-of-the-art results have been achieved in terms of mass detection accuracy. Satisfactory mass malignancy classification result has also been obtained on our in-house dataset.

Human fingerprints are detailed and nearly unique markers of human identity. Such a unique and stable fingerprint is also left on each acquired image. It can reveal how an image was degraded during the image acquisition procedure and thus is closely related to the quality of an image. In this work, we propose a new no-reference image quality assessment (NR-IQA) approach called domain-aware IQA (DA-IQA), which for the first time introduces the concept of domain fingerprint to the NR-IQA field. The domain fingerprint of an image is learned from image collections of different degradations and then used as the unique characteristics to identify the degradation sources and assess the quality of the image. To this end, we design a new domain-aware architecture, which enables simultaneous determination of both the distortion sources and the quality of an image. With the distortion in an image better characterized, the image quality can be more accurately assessed, as verified by extensive experiments, which show that the proposed DA-IQA performs better than almost all the compared state-of-the-art NR-IQA methods.

Breast Ultrasound (BUS) imaging has been recognized as an essential imaging modality for breast masses classification in China. Current deep learning (DL) based solutions for BUS classification seek to feed ultrasound (US) images into deep convolutional neural networks (CNNs), to learn a hierarchical combination of features for discriminating malignant and benign masses. One existing problem in current DL-based BUS classification was the lack of spatial and channel-wise features weighting, which inevitably allow interference from redundant features and low sensitivity. In this study, we aim to incorporate the instructive information provided by breast imaging reporting and data system (BI-RADS) within DL-based classification. A novel DL-based BI-RADS Vector-Attention Network (BVA Net) that trains with both texture information and decoded information from BI-RADS stratifications was proposed for the task. Three baseline models, pre-trained DenseNet-121, ResNet-50 and Residual-Attention Network (RA Net) were included for comparison. Experiments were conducted on a large scale private main dataset and two public datasets, UDIAT and BUSI. On the main dataset, BVA Net outperformed other models, in terms of AUC (area under the receiver operating curve, 0.908), ACC (accuracy, 0.865), sensitivity (0.812) and precision (0.795). BVA Net also achieved the high AUC (0.87 and 0.882) and ACC (0.859 and 0.843), on UDIAT and BUSI. Moreover, we proposed a method that integrates both BVA Net binary classification and BI-RADS stratification estimation, called integrated classification. The introduction of integrated classification helped improving the overall sensitivity while maintaining a high specificity.

The acquisition of large-scale medical image data, necessary for training machine learning algorithms, is hampered by associated expert-driven annotation costs. Mining hospital archives can address this problem, but labels often incomplete or noisy, e.g., 50% of the lesions in DeepLesion are left unlabeled. Thus, effective label harvesting methods are critical. This is the goal of our work, where we introduce Lesion-Harvester-a powerful system to harvest missing annotations from lesion datasets at high precision. Accepting the need for some degree of expert labor, we use a small fully-labeled image subset to intelligently mine annotations from the remainder. To do this, we chain together a highly sensitive lesion proposal generator (LPG) and a very selective lesion proposal classifier (LPC). Using a new hard negative suppression loss, the resulting harvested and hard-negative proposals are then employed to iteratively finetune our LPG. While our framework is generic, we optimize our performance by proposing a new 3D contextual LPG and by using a global-local multi-view LPC. Experiments on DeepLesion demonstrate that Lesion-Harvester can discover an additional 9,805 lesions at a precision of 90%. We publicly release the harvested lesions, along with a new test set of completely annotated DeepLesion volumes. We also present a pseudo 3D IoU evaluation metric that corresponds much better to the real 3D IoU than current DeepLesion evaluation metrics. To quantify the downstream benefits of Lesion-Harvester we show that augmenting the DeepLesion annotations with our harvested lesions allows state-of-the-art detectors to boost their average precision by 7 to 10%.

Accurate segmentation of anatomical structures is vital for medical image analysis. The state-of-the-art accuracy is typically achieved by supervised learning methods, where gathering the requisite expert-labeled image annotations in a scalable manner remains a main obstacle. Therefore, annotation-efficient methods that permit to produce accurate anatomical structure segmentation are highly desirable. In this work, we present Contour Transformer Network (CTN), a one-shot anatomy segmentation method with a naturally built-in human-in-the-loop mechanism. We formulate anatomy segmentation as a contour evolution process and model the evolution behavior by graph convolutional networks (GCNs). Training the CTN model requires only one labeled image exemplar and leverages additional unlabeled data through newly introduced loss functions that measure the global shape and appearance consistency of contours. On segmentation tasks of four different anatomies, we demonstrate that our one-shot learning method significantly outperforms non-learning-based methods and performs competitively to the state-of-the-art fully supervised deep learning methods. With minimal human-in-the-loop editing feedback, the segmentation performance can be further improved to surpass the fully supervised methods.

Large-scale datasets with high-quality labels are desired for training accurate deep learning models. However, due to the annotation cost, datasets in medical imaging are often either partially-labeled or small. For example, DeepLesion is such a large-scale CT image dataset with lesions of various types, but it also has many unlabeled lesions (missing annotations). When training a lesion detector on a partially-labeled dataset, the missing annotations will generate incorrect negative signals and degrade the performance. Besides DeepLesion, there are several small single-type datasets, such as LUNA for lung nodules and LiTS for liver tumors. These datasets have heterogeneous label scopes, i.e., different lesion types are labeled in different datasets with other types ignored. In this work, we aim to develop a universal lesion detection algorithm to detect a variety of lesions. The problem of heterogeneous and partial labels is tackled. First, we build a simple yet effective lesion detection framework named Lesion ENSemble (LENS). LENS can efficiently learn from multiple heterogeneous lesion datasets in a multi-task fashion and leverage their synergy by proposal fusion. Next, we propose strategies to mine missing annotations from partially-labeled datasets by exploiting clinical prior knowledge and cross-dataset knowledge transfer. Finally, we train our framework on four public lesion datasets and evaluate it on 800 manually-labeled sub-volumes in DeepLesion. Our method brings a relative improvement of 49% compared to the current state-of-the-art approach in the metric of average sensitivity. We have publicly released our manual 3D annotations of DeepLesion online.11https://github.com/viggin/DeepLesion_manual_test_set

Owing to its unique literal and aesthetical characteristics, automatic generation of Chinese poetry is still challenging in Artificial Intelligence, which can hardly be straightforwardly realized by end-to-end methods. In this paper, we propose a novel iterative polishing framework for highly qualified Chinese poetry generation. In the first stage, an encoder-decoder structure is utilized to generate a poem draft. Afterwards, our proposed Quality-Aware Masked Language Model (QA-MLM) is employed to polish the draft towards higher quality in terms of linguistics and literalness. Based on a multi-task learning scheme, QA-MLM is able to determine whether polishing is needed based on the poem draft. Furthermore, QA-MLM is able to localize improper characters of the poem draft and substitute with newly predicted ones accordingly. Benefited from the masked language model structure, QA-MLM incorporates global context information into the polishing process, which can obtain more appropriate polishing results than the unidirectional sequential decoding. Moreover, the iterative polishing process will be terminated automatically when QA-MLM regards the processed poem as a qualified one. Both human and automatic evaluation have been conducted, and the results demonstrate that our approach is effective to improve the performance of encoder-decoder structure.

Social connections play a vital role in improving the performance of recommendation systems (RS). However, incorporating social information into RS is challenging. Most existing models usually consider social influences in a given session, ignoring that both users? preferences and their friends? influences are evolving. Moreover, in real world, social relations are sparse. Modeling dynamic influences and alleviating data sparsity is of great importance.

OAR segmentation is a critical step in radiotherapy of head and neck (H&N) cancer, where inconsistencies across radiation oncologists and prohibitive labor costs motivate automated approaches. However, leading methods using standard fully convolutional network workflows that are challenged when the number of OARs becomes large, e.g. > 40. For such scenarios, insights can be gained from the stratification approaches seen in manual clinical OAR delineation. This is the goal of our work, where we introduce stratified organ at risk segmentation (SOARS), an approach that stratifies OARs into anchor, mid-level, and small & hard (S&H) categories. SOARS stratifies across two dimensions. The first dimension is that distinct processing pipelines are used for each OAR category. In particular, inspired by clinical practices, anchor OARs are used to guide the mid-level and S&H categories. The second dimension is that distinct network architectures are used to manage the significant contrast, size, and anatomy variations between different OARs. We use differentiable neural architecture search (NAS), allowing the network to choose among 2D, 3D or Pseudo-3D convolutions. Extensive 4-fold cross-validation on 142 H&N cancer patients with 42 manually labeled OARs, the most comprehensive OAR dataset to date, demonstrates that both pipeline- and NAS-stratification significantly improves quantitative performance over the state-of-the-art (from 69.52% to 73.68% in absolute Dice scores). Thus, SOARS provides a powerful and principled means to manage the highly complex segmentation space of OARs.

Mammography plays an essential role in early detection of breast cancer. Interpreting mammography is a professional task that requires well-trained radiologists with longtime clinical experience. In this paper, we present MABEL, an artificial intelligence-powered system to assist doctors for breast cancer screening and diagnosis in mammograms, in order to reduce their workloads and accelerate the diagnostic process. Our system smoothly integrates our upgraded lesion identification models, provides a doctor-oriented annotation tool and web interface, and can communicate with Picture Archiving and Communication System (PACS) in our collaborative hospital. Our lesion identification performance is evaluated on both public and in-house datasets, in which mass detection has achieved state-of-the-art accuracy in the single-view manner. The overall high satisfaction from doctors of our system is also demonstrated.

Recently, the speaker clustering model based on aggregation hierarchy cluster (AHC) is a common method to solve two main problems: no preset category number clustering and fix category number clustering. In general, model takes features like i-vectors as input of probability and linear discriminant analysis model (PLDA) aims to form the distance matric in long voice application scenario, and then clustering results are obtained through the clustering model. However, traditional speaker clustering method based on AHC has the shortcomings of long-time running and remains sensitive to environment noise. In this paper, we propose a novel speaker clustering method based on Mutual Information (MI) and a non-linear model with discrete variable, which under the enlightenment of Tied Variational Autoencoder (TVAE), to enhance the robustness against noise. The proposed method named Discrete Tied Variational Autoencoder (DTVAE) which shortens the elapsed time substantially. With experience results, it outperforms the general model and yields a relative Accuracy (ACC) improvement and significant time reduction.

In this work, we propose Flow-TTS, a non-autoregressive end-to-end neural TTS model based on generative flow. Unlike other non-autoregressive models, Flow-TTS can achieve high-quality speech generation by using a single feed-forward network. To our knowledge, Flow-TTS is the first TTS model utilizing flow in spectrogram generation network and the first non-autoregssive model which jointly learns the alignment and spectrogram generation through a single network. Experiments on LJSpeech show that the speech quality of Flow-TTS heavily approaches that of human and is even better than that of autoregressive model Tacotron 2 (outperforms Tacotron 2 with a gap of 0.09 in MOS). Meanwhile, the inference speed of Flow-TTS is about 23 times speed-up over Tacotron 2, which is comparable to FastSpeech.1

This paper leverages the graph-to-sequence method in neural text-to-speech (GraphTTS), which maps the graph embedding of the input sequence to spectrograms. The graphical inputs consist of node and edge representations constructed from input texts. The encoding of these graphical inputs incorporates syntax information by a GNN encoder module. Besides, applying the encoder of GraphTTS as a graph auxiliary encoder (GAE) can analyse prosody information from the semantic structure of texts. This can remove the manual selection of reference audios process and makes prosody modelling an end-to-end procedure. Experimental analysis shows that GraphTTS outperforms the state-of-the-art sequence-to-sequence models by 0.24 in Mean Opinion Score (MOS). GAE can adjust the pause, ventilation and tones of synthesised audios automatically. This experimental conclusion may give some inspiration to researchers working on improving speech synthesis prosody.

Targeting at both high efficiency and performance, we propose AlignTTS to predict the mel-spectrum in parallel. AlignTTS is based on a Feed-Forward Transformer which generates mel-spectrum from a sequence of characters, and the duration of each character is determined by a duration predictor. Instead of adopting the attention mechanism in Transformer TTS to align text to mel-spectrum, the alignment loss is presented to consider all possible alignments in training by use of dynamic programming. Experiments on the LJSpeech dataset show that our model achieves not only state-of-the-art performance which outperforms Transformer TTS by 0.03 in mean option score (MOS), but also a high efficiency which is more than 50 times faster than real-time.

A number of methods based on deep learning have been applied to medical image segmentation and have achieved state-of-the-art performance. Due to the importance of chest x-ray data in studying COVID-19, there is a demand for state-of-the-art models capable of precisely segmenting soft tissue on the chest x-rays. The dataset for exploring best segmentation model is from Montgomery and Shenzhen hospital which had opened in 2014. The most famous technique is U-Net which has been used to many medical datasets including the Chest X-rays. However, most variant U-Nets mainly focus on extraction of contextual information and skip connections. There is still a large space for improving extraction of spatial features. In this paper, we propose a dual encoder fusion U-Net framework for Chest X-rays based on Inception Convolutional Neural Network with dilation, Densely Connected Recurrent Convolutional Neural Network, which is named DEFU-Net. The densely connected recurrent path extends the network deeper for facilitating contextual feature extraction. In order to increase the width of network and enrich representation of features, the inception blocks with dilation are adopted. The inception blocks can capture globally and locally spatial information from various receptive fields. At the same time, the two paths are fused by summing features, thus preserving the contextual and spatial information for decoding part. This multi-learning-scale model is benefiting in Chest X-ray dataset from two different manufacturers (Montgomery and Shenzhen hospital). The DEFU-Net achieves the better performance than basic U-Net, residual U-Net, BCDU-Net, R2U-Net and attention R2U-Net. This model has proved the feasibility for mixed dataset and approaches state-of-the-art. The source code for this proposed framework is public https://github.com/uceclz0/DEFU-Net.

Due to the improvement of Language Modeling, the emerging NLP assistant tools aiming for text generation greatly reduce the human workload on writing documents. However, the generation of legal text faces greater challenges than ordinary texts because of its high requirement for keeping logic reasonable, which can not be guaranteed by Language Modeling right now. To generate reasonable legal documents, we propose a novel method CoLMQA, which (1) combines Language Modeling and Question Answering, (2) generates text with slots by Language Modeling, and (3) fills the slots by our proposed Question Answering method named Transformer-based Key-Value Memory Networks. In CoLMQA, the slots represent the text part that needs to be highly constrained by logic, such as the name of the law and the number of the law article. And the Question Answering fills the slots in context with the help of Legal Knowledge Base to keep logic reasonable. The experiment verifies the quality of legal documents generated by CoLMQA, surpassing the documents generated by pure Language Modeling.

End-to-end Spoken Language Understanding (SLU) models are made increasingly large and complex to achieve the state-ofthe-art accuracy.However, the increased complexity of a model can also introduce high risk of over-fitting, which is a major challenge in SLU tasks due to the limitation of available data.In this paper, we propose an attention-based SLU model together with three encoder enhancement strategies to overcome data sparsity challenge.The first strategy focuses on the transferlearning approach to improve feature extraction capability of the encoder.It is implemented by pre-training the encoder component with a quantity of Automatic Speech Recognition annotated data relying on the standard Transformer architecture and then fine-tuning the SLU model with a small amount of target labelled data.The second strategy adopts multitask learning strategy, the SLU model integrates the speech recognition model by sharing the same underlying encoder, such that improving robustness and generalization ability.The third strategy, learning from Component Fusion (CF) idea, involves a Bidirectional Encoder Representation from Transformer (BERT) model and aims to boost the capability of the decoder with an auxiliary network.It hence reduces the risk of over-fitting and augments the ability of the underlying encoder, indirectly.Experiments on the FluentAI dataset show that cross-language transfer learning and multi-task strategies have been improved by up to 4.52% and 3.89% respectively, compared to the baseline.

State-of-the-art speaker verification models are based on deep learning techniques, which heavily depend on the handdesigned neural architectures from experts or engineers. We borrow the idea of neural architecture search(NAS) for the textindependent speaker verification task. As NAS can learn deep network structures automatically, we introduce the NAS conception into the well-known x-vector network. Furthermore, this paper proposes an evolutionary algorithm enhanced neural architecture search method called Auto-Vector to automatically discover promising networks for the speaker verification task. The experimental results demonstrate our NAS-based model outperforms state-of-the-art speaker verification models.

In this paper, we propose a real-time robot-based auxiliary system for risk evaluation of COVID-19 infection.It combines real-time speech recognition, temperature measurement, keyword detection, cough detection and other functions in order to convert live audio into actionable structured data to achieve the COVID-19 infection risk assessment function.In order to better evaluate the COVID-19 infection, we propose an end-to-end method for cough detection and classification for our proposed system.It is based on real conversation data from human-robot, which processes speech signals to detect cough and classifies it if detected.The structure of our model are maintained concise to be implemented for real-time applications.And we further embed this entire auxiliary diagnostic system in the robot and it is placed in the communities, hospitals and supermarkets to support COVID-19 testing.The system can be further leveraged within a business rules engine, thus serving as a foundation for real-time supervision and assistance applications.Our model utilizes a pretrained, robust training environment that allows for efficient creation and customization of customer-specific health states.

Recent neural speech synthesis systems have gradually focused on the control of prosody to improve the quality of synthesized speech, but they rarely consider the variability of prosody and the correlation between prosody and semantics together.In this paper, a prosody learning mechanism is proposed to model the prosody of speech based on TTS system, where the prosody information of speech is extracted from the melspectrum by a prosody learner and combined with the phoneme sequence to reconstruct the mel-spectrum.Meanwhile, the sematic features of text from the pre-trained language model is introduced to improve the prosody prediction results.In addition, a novel self-attention structure, named as local attention, is proposed to lift this restriction of input text length, where the relative position information of the sequence is modeled by the relative position matrices so that the position encodings is no longer needed.Experiments on English and Mandarin show that speech with more satisfactory prosody has obtained in our model.Especially in Mandarin synthesis, our proposed model outperforms baseline model with a MOS gap of 0.08, and the overall naturalness of the synthesized speech has been significantly improved.

Voice activity detection (VAD) makes a distinction between speech and non-speech and its performance is of crucial importance for speech based services.Recently, deep neural network (DNN)-based VADs have achieved better performance than conventional signal processing methods.The existed DNNbased models always handcrafted a fixed window to make use of the contextual speech information to improve the performance of VAD.However, the fixed window of contextual speech information can't handle various unpredictable noise environments and highlight the critical speech information to VAD task.In order to solve this problem, this paper proposed an adaptive multiple receptive-field attention neural network, called MLNET, to finish VAD task.The MLNET leveraged multi-branches to extract multiple contextual speech information and investigated an effective attention block to weight the most crucial parts of the context for final classification.Experiments in real-world scenarios demonstrated that the proposed MLNET-based model outperformed other baselines.

Emotion recognition in conversation (ERC) is an important topic for developing empathetic machines in a variety of areas including social opinion mining, health-care and so on.In this paper, we propose a method to model ERC task as sequence tagging where a Conditional Random Field (CRF) layer is leveraged to learn the emotional consistency in the conversation.We employ LSTM-based encoders that capture self and inter-speaker dependency of interlocutors to generate contextualized utterance representations which are fed into the CRF layer.For capturing long-range global context, we use a multi-layer Transformer encoder to enhance the LSTM-based encoder.Experiments show that our method benefits from modeling the emotional consistency and outperforms the current state-of-theart methods on multiple emotion classification datasets.

Internet insurance products are apparently different from traditional e-commerce goods for their complexity, low purchasing frequency, etc. So, cold start problem is even worse. In traditional e-commerce field, several cross-domain recommendation (CDR) methods have been studied to infer preferences of cold start users based on their preferences in other domains. However, these CDR methods couldn't be applied into insurance domain directly due to product complexity. In this paper, we propose a Deep Cross-Domain Insurance Recommendation System (DCDIR) for cold start users. Specifically, we first learn more effective user and item latent features in both domains. In target domain, given the complexity of insurance products, we design a meta-path based method over insurance product knowledge graph. In source domain, we employ GRU to model users' dynamic interests. Then we learn a feature mapping function by multi-layer perceptions. We apply DCDIR on our company's dataset, and show DCDIR significantly outperforms the state-of-the-art solutions.

Internet is changing the world, adapting to the trend of internet sales will bring revenue to traditional insurance companies. Online insurance is still in its early stages of development, where cold start problem (prospective customer) is one of the greatest challenges. In traditional e-commerce field, several cross-domain recommendation (CDR) methods have been studied to infer preferences of cold start users based on their preferences in other domains. However, these CDR methods couldn't be applied to insurance domain directly due to the domain's specific properties. In this paper, we propose a novel framework called a Heterogeneous information network based Cross Domain Insurance Recommendation (HCDIR) system for cold start users. Specifically, we first try to learn more effective user and item latent features in both source and target domains. In source domain, we employ gated recurrent unit (GRU) to module users' dynamic interests. In target domain, given the complexity of insurance products and the data sparsity problem, we construct an insurance heterogeneous information network (IHIN) based on data from PingAn Jinguanjia, the IHIN connects users, agents, insurance products and insurance product properties together, giving us richer information. Then we employ three-level (relational, node, and semantic) attention aggregations to get user and insurance product representations. After obtaining latent features of overlapping users, a feature mapping between the two domains is learned by multi-layer perceptron (MLP). We apply HCDIR on Jinguanjia dataset, and show HCDIR significantly outperforms the state-of-the-art solutions.

This paper presents a novel approach to train GANs for discrete sequence generation without resorting to an explicit neural network as the discriminator. We show that when an alternative mini-max optimization procedure is performed for the value function where a closed form solution for the discriminator exists in the maximization step, it is equivalent to directly optimizing the Jenson-Shannon divergence (JSD) between the generator’s distribution and the empirical distribution over the training data without sampling from the generator, thus optimizing the JSD becomes computationally tractable to train the generator that generates sequences of discrete data. Extensive experiments on synthetic data and real-world tasks demonstrate significant improvements over existing methods to train GANs that generate discrete sequences.

Unsegmented text recognition is a crucial component in financial document processing systems. Financial text materials, such as receipts, transcripts, identification documents, etc. often involve critical personal information. In many circumstances, these data reside in protected servers of different institutions and must not be transferred beyond the institutional firewall. The emerging technology of Federated Learning (FL) provides a data-secure way of uniting isolated datasets in model training. Using the FL framework, text recognition models can be trained with larger collection of image samples. In previous works, federated text recognition models only deal with single-character images and alpha-numeric corpus. Such models are not competent in industrial applications, especially in Chinese text recognition problems. In this paper, we apply federated learning with a deep convolutional network to perform variable-length text string recognition with a large corpus. In our experiments, we compared two prevalent federated learning frameworks, namely, Tensorflow Federated and PySyft. Results show that federated text recognition models can achieve similar or even higher accuracy than models trained on deep learning framework. On a 5-client distributed dataset, the best character accuracy is achieved by TFF at 49.20%. Extensive experiments are also conducted to evaluate the effect of distributed data storage over the performance of trained models. TFF again achieved a maximum character precision of 54.33% with non-distributed dataset.

Seasonal influenza epidemics which annually cause substantial diseases and deaths in high-risk population groups are a major public health concern around the world. Considering the hysteresis of traditional flu surveillance systems, this work aims to present a methodology capable of forecasting influenza activity of a city in China precisely 1 week ahead of the official publication. To that end, exogenous information collected from different sources were separately tested with historical influenza-like illness reports for the ability of detecting influenza activity, including climate surveillance, Internet users’ search activity, twitter and health inquiry on an online health consultation platform. Moreover, an ensemble model combining a time series analysis model and a tree boosting model based on those multisource data was applied to improve the accuracy and generalizability of influenza forecasting, in which a model fusion method based on the Kalman Filter was proposed. The validation experiments in this work were performed on the influenza-like illness reports collected from Chongqing city over 4 influenza seasons within 2014–2017. The results show that the proposed model outperformed other tested models by not only taking the periodic law of influenza into consideration but also incorporating information from diverse data sources. The mean absolute percentage error of the validation data set decreased to about 10%. This work provides a viable suggestion for improving the influenza activity forecasting of a city at its early stage.

We describe a multitarget speaker detection system that provides a robust way to classify the utterance of a speaker in noisy environments. The multitarget detection problem is known to be much more difficult to tackle than single target speaker verification tasks, especially when the target set is large and the data is corrupted by noise. In this work we aim to improve the performance of our multitarget speaker detection system in real-world settings, where complicated background noise and unpredictable speaker behavior are present. We make three major improvements that contribute to our goal. First, we discover an effective noise-filtering method using GMM-based voice activity detector followed by unsupervised bottom-up clustering. Second, we incorporate a Highway-LSTM network to estimate posterior distributions of senones, replacing the traditional GMM-UBM with senone posteriors. Finally, we apply a self-adaptive approach on the classifier back-end so that our PLDA parameters and S-normalization subsets can be updated online.

Social network services have become important and efficient platforms for users to share all kinds of information. The capability to monitor user-generated information and detect bursts from information diffusions in these social networks brings value to a wide range of real-life applications, such as viral marketing. However, in reality, as a third party, there is always a cost for gathering information from each user or so-called social network sensor. The question then arises how to select a budgeted set of social network sensors to form the data stream for burst detection without compromising the detection performance. In this article, we present a general sensor selection solution for different burst detection approaches. We formulate this problem as a constraint satisfaction problem that has high computational complexity. To reduce the computational cost, we first reduce most of the constraints by making use of the fact that bursty cascades are rare among the whole population. We then transform the problem into an Linear Programming (LP) problem. Furthermore, we use the sub-gradient method instead of the standard simplex method or interior-point method to solve the LP problem, which makes it possible for our solution to scale up to large social networks. Evaluating our solution on millions of real information cascades, we demonstrate both the effectiveness and efficiency of our approach.

User identity linkage across social platforms is an important problem of great research challenge and practical value. In real applications, the task often assumes an extra degree of difficulty by requiring linkage across multiple platforms. While pair-wise user linkage between two platforms, which has been the focus of most existing solutions, provides reasonably convincing linkage, the result depends by nature on the order of platform pairs in execution with no theoretical guarantee on its stability. In this paper, we explore a new concept of ``Latent User Space'' to more naturally model the relationship between the underlying real users and their observed projections onto the varied social platforms, such that the more similar the real users, the closer their profiles in the latent user space. We propose two effective algorithms, a batch model(ULink) and an online model(ULink-On), based on latent user space modelling. Two simple yet effective optimization methods are used for optimizing objective function: the first one based on the constrained concave-convex procedure(CCCP) and the second on accelerated proximal gradient. To our best knowledge, this is the first work to propose a unified framework to address the following two important aspects of the multi-platform user identity linkage problem --- (I) the platform multiplicity and (II) online data generation. We present experimental evaluations on real-world data sets for not only traditional pairwise-platform linkage but also multi-platform linkage. The results demonstrate the superiority of our proposed method over the state-of-the-art ones.

Context has been playing an increasingly important role to improve the object detection performance. In this paper we propose an effective representation, Multi-Order Contextual co-Occurrence (MOCO), to implicitly model the high level context using solely detection responses from a baseline object detector. The so-called (1st-order) context feature is computed as a set of randomized binary comparisons on the response map of the baseline object detector. The statistics of the 1st-order binary context features are further calculated to construct a high order co-occurrence descriptor. Combining the MOCO feature with the original image feature, we can evolve the baseline object detector to a stronger context aware detector. With the updated detector, we can continue the evolution till the contextual improvements saturate. Using the successful deformable-part-model detector [13] as the baseline detector, we test the proposed MOCO evolution framework on the PASCAL VOC 2007 dataset [8] and Caltech pedestrian dataset [7]: The proposed MOCO detector outperforms all known state-of-the-art approaches, contextually boosting deformable part models (ver. 5) [13] by 3.3% in mean average precision on the PASCAL 2007 dataset. For the Caltech pedestrian dataset, our method further reduces the log-average miss rate from 48% to 46% and the miss rate at 1 FPPI from 25% to 23%, compared with the best prior art [6].

Recognizing continuous action composition in human behavior is an important and yet challenging problem. In this paper we tackle the task by developing both reliable image features and classification algorithms. For image features, we introduce the Embedded Optical Flow (EOF) feature based on embedding optical flow using Locality-constrained Linear Coding with weighted average pooling. The EOF feature is histogram-like but presents excellent linear separability. For classification, we propose the Continuous Linear Dynamic System (CLDS) framework that consists of two sets of Linear Dynamic System (LDS) models, one to model the dynamics of individual actions and the other to model the transition between actions. The inference process estimates the best decomposition of the whole sequence into continuous alternating between human actions and action transitions. In this way, both action type and action boundary can be accurately recognized. Extensive experiments demonstrate the effectiveness and efficiency of the proposed EOF feature and CLDS algorithm.

Pedestrian detection from images is an important and yet challenging task. The conventional methods usually identify human figures using image features inside the local regions. In this paper we present that, besides the local features, context cues in the neighborhood provide important constraints that are not yet well utilized. We propose a framework to incorporate the context constraints for detection. First, we combine the local window with neighborhood windows to construct a multi-scale image context descriptor, designed to represent the contextual cues in spatial, scaling, and color spaces. Second, we develop an iterative classification algorithm called contextual boost. At each iteration, the classifier responses from the previous iteration across the neighborhood and multiple image scales, called classification context, are incorporated as additional features to learn a new classifier. The number of iterations is determined in the training process when the error rate converges. Since the classification context incorporates contextual cues from the neighborhood, through iterations it implicitly propagates to greater areas and thus provides more global constraints. We evaluate our method on the Caltech benchmark dataset [11]. The results confirm the advantages of the proposed framework. Compared with state of the arts, our method reduces the miss rate from 29% by [30] to 25% at 1 false positive per image (FPPI).

This paper introduces a probabilistic graphical model for continuous action recognition with two novel components: substructure transition model and discriminative boundary model. The first component encodes the sparse and global temporal transition prior between action primitives in state-space model to handle the large spatial-temporal variations within an action class. The second component enforces the action duration constraint in a discriminative way to locate the transition boundaries between actions more accurately. The two components are integrated into a unified graphical structure to enable effective training and inference. Our comprehensive experimental results on both public and in-house datasets show that, with the capability to incorporate additional information that had not been explicitly or efficiently modeled by previous methods, our proposed algorithm achieved significantly improved performance for continuous action recognition.

Content-aware image retargeting has attracted a lot of interests recently. The key and most challenging issue for this task is how to balance the tradeoff between preserving the important contents and minimizing the visual distortions on the consistency of the image structure. In this paper we present a novel filtering-based technique to tackle this issue, called ”importance filtering”. Specifically, we first filter the image saliency guided by the image itself to achieve a structure-consistent importance map. We then use the pixel importance as the key constraint to compute the gradient map of pixel shifts from the original resolution to the target. Finally, we integrate the shift gradient across the image using a weighted filter to construct a smooth shift map and render the target image. The weight is again controlled by the pixel importance. The two filtering processes enforce to maintain the structural consistency and yet preserve the important contents in the target image. Furthermore, the simple nature of filter operations allows highly efficient implementation for real-time applications and easy extension to video retargeting, as the structural constraints from the original image naturally convey the temporal coherence between frames. The effectiveness and efficiency of our importance filtering algorithm are confirmed in extensive experiments.

We present a novel theory for characterizing defocus blurs in multi-perspective cameras such as catadioptric mirrors. Our approach studies how multi-perspective ray geometry transforms under the thin lens. We first use the General Linear Cameras (GLCs) to approximate the incident multi-perspective rays to the lens and then apply a Thin Lens Operator (TLO) to map an incident GLC to the exit GLC. To study defocus blurs caused by the GLC rays, we further introduce a new Ray Spread Function (RSF) model analogous the Point Spread Function (PSF). While PSF models defocus blurs caused by a 3D scene point, RSF models blurs spread by rays. We derive closed form RSFs for incident GLC rays, and we show that for catadioptric cameras with a circular aperture, the RSF can be effectively approximated as a single or mixtures of elliptic-shaped kernels. We apply our method for predicting defocus blurs on commonly used catadioptric cameras and for reducing de-focus blurs in catadioptric projections. Experiments on synthetic and real data demonstrate the accuracy and general applicability of our approach.

Non-negative matrix factorization (NMF) has previously been shown to be a useful decomposition tool for multivariate data. Non-negative bases allow strictly additive combinations which have been shown to be part-based as well as relatively sparse. We pursue a discriminative decomposition by coupling NMF objective with a maximum margin classifier, specifically a support vector machine (SVM). Conversely, we propose an NMF based regularizer for SVM. We formulate the joint update equations and propose a new method which identifies the decomposition as well as the classification parameters. We present classification results on synthetic as well as real datasets.

We present a catadioptric projector analogous to a catadioptric camera by combining a commodity digital projector with additional optical units. We show that, by using specially shaped reflectors/refractors, catadioptric projectors can offer an unprecedented level of flexibility in aspect ratio, size, and field of view. We also present efficient algorithms to reduce projection artifacts in catadioptric projectors, such as distortions, scattering, and defocusing. Instead of recovering the reflector/refractor geometry, our approach directly models the light transport between the projector and the viewpoint using the light transport matrix (LTM). We show how to efficiently approximate the pseudo inverse of the LTM and apply it to find the optimal input image that produces least projection distortions. Furthermore, we present a projection defocus analysis for reflector and thin refractor based catadioptric projectors. We show that defocus blur can be interpreted as spatially-varying Gaussian blurs on the input image. We then measure the kernels directly from the LTM and apply deconvolution to optimize the input image. We demonstrate the practical uses of catadioptric projectors in panoramic and omni-directional projections. Our new system achieves much wider field-of-view projection while maintaining sharpness and low geometric and photometric distortions.

Active Appearance Models (AAMs) have been popularly used to represent the appearance and shape variations of human faces. Fitting an AAM to images recovers the face pose as well as its deformable shape and varying appearance. Successful fitting requires that the AAM is sufficiently generic such that it covers all possible facial appearances and shapes in the images. Such a generic AAM is often difficult to be obtained in practice, especially when the image quality is low or when occlusion occurs. To achieve robust AAM fitting under such circumstances, this paper proposes to incorporate the disparity data obtained from a stereo camera with the image fitting process. We develop an iterative multi-level algorithm that combines efficient AAM fitting to 2D images and robust 3D shape alignment to disparity data. Experiments on tracking faces in low-resolution images captured from meeting scenarios show that the proposed method achieves better performance than the original 2D AAM fitting algorithm. We also demonstrate an application of the proposed method to a facial expression recognition task.

Many natural objects vary the shapes as linear combinations of certain bases. The measurement of such deformable shapes is coupling of rigid similarity transformations between the objects and the measuring systems and non-rigid deformations controlled by the linear bases. Thus registration and modeling of deformable shapes are coupled problems, where registration is to compute the rigid transformations and modeling is to construct the linear bases. The previous methods [3, 2] separate the solution into two steps. The first step registers the measurements regarding the shapes as rigid and the deformations as random noise. The second step constructs the linear model using the registered shapes. Since the deformable shapes do not vary randomly but are constrained by the underlying model, such separate steps result in registration biased by nonrigid deformations and shape models involving improper rigid transformations. We for the first time present this bias problem and formulate that, the coupled registration and modeling problems are essentially a single factorization problem and thus require a simultaneous solution. We then propose the Direct Factorization method that extends a structure from motion method [16]. It yields a linear closedform solution that simultaneously registers the deformable shapes at arbitrary dimensions (2D \to 2D, 3D \to 3D, . . .) and constructs the linear bases. The accuracy and robustness of the proposed approach are demonstrated quantitatively on synthetic data and qualitatively on real shapes.

We propose a system that is capable of detailed analysis of eye region images in terms of the position of the iris, degree of eyelid opening, and the shape, complexity, and texture of the eyelids. The system uses a generative eye region model that parameterizes the fine structure and motion of an eye. The structure parameters represent structural individuality of the eye, including the size and color of the iris, the width, boldness, and complexity of the eyelids, the width of the bulge below the eye, and the width of the illumination reflection on the bulge. The motion parameters represent movement of the eye, including the up-down position of the upper and lower eyelids and the 2D position of the iris. The system first registers the eye model to the input in a particular frame and individualizes it by adjusting the structure parameters. The system then tracks motion of the eye by estimating the motion parameters across the entire image sequence. Combined with image stabilization to compensate for appearance changes due to head motion, the system achieves accurate registration and motion recovery of eyes.

In this paper, we study the relationship between multi-view active appearance model (AAM) fitting and camera calibration. In the first part of the paper we propose an algorithm to calibrate the relative orientation of a set of N > 1 cameras by fitting an AAM to sets of N images. In essence, we use the human face as a (non-rigid) calibration grid. Our algorithm calibrates a set of 2 /spl times/ 3 weak-perspective camera projection matrices, protections of the world coordinate system origin into the images, depths of the world coordinate system origin, and focal lengths. We demonstrate that the performance of this algorithm is comparable to a standard algorithm using a calibration grid. In the second part of the paper, we show how calibrating the cameras improves tile performance of multi-view AAM fitting.

Reconstruction of 3D structures from uncalibrated image sequences has a wealthy history. Most work has been focused on rigid objects or static scenes. This paper studies the problem of perspective reconstruction of deformable structures such as dynamic scenes from an uncalibrated image sequence. The task requires decomposing the image measurements into a composition of three factors: 3D deformable structures, rigid rotations and translations, and intrinsic camera parameters. We develop a factorization algorithm that consists of two steps. In the first step we recover the protective depths iteratively using the sub-space constraints embedded in the image measurements of the deformable structures. In the second step, we scale the image measurements by the reconstructed projective depths. We then extend the linear closed-form solution for weak-perspective reconstruction by J. Xiao, et al. (2004) to factorize the scaled measurements and simultaneously reconstruct the deformable shapes and underlying shape model, the rigid motions, and the varying camera parameters such as focal lengths. The accuracy and robustness of the proposed method is demonstrated quantitatively on synthetic data and qualitatively on real image sequences.

Active Appearance Models (AAMs) are a well studied 2D deformable model. One recently proposed extension of AAMs to multiple images is the Coupled-View AAM. Coupled-View AAMs model the 2D shape and appearance of a face in two or more views simultaneously. The major limitation of Coupled-View AAMs, however, is that they are specific to a particular set of cameras, both in geometry and the photometric responses. In this paper, we describe how a single AAM can be fit to multiple images, captured simultaneously by cameras with arbitrary geometry and response functions. Our algorithm retains the major benefits of Coupled-View AAMs: the integration of information from multiple images into a single model, and improved fitting robustness. 1