The introduction of diffusion models has brought significant advances to the field of audio-driven talking head generation. However, the extremely slow inference speed severely limits the practical implementation of diffusion-based talking head generation models. In this study, we propose READ, a real-time diffusion-transformer-based talking head generation framework. Our approach first learns a spatiotemporal highly compressed video latent space via a temporal VAE, significantly reducing the token count to accelerate generation. To achieve better audio-visual alignment within this compressed latent space, a pre-trained Speech Autoencoder (SpeechAE) is proposed to generate temporally compressed speech latent codes corresponding to the video latent space. These latent representations are then modeled by a carefully designed Audio-to-Video Diffusion Transformer (A2V-DiT) backbone for efficient talking head synthesis. Furthermore, to ensure temporal consistency and accelerated inference in extended generation, we propose a novel asynchronous noise scheduler (ANS) for both the training and inference processes of our framework. The ANS leverages asynchronous add-noise and asynchronous motion-guided generation in the latent space, ensuring consistency in generated video clips. Experimental results demonstrate that READ outperforms state-of-the-art methods by generating competitive talking head videos with significantly reduced runtime, achieving an optimal balance between quality and speed while maintaining robust metric stability in long-time generation.

3D Gaussian Splatting (3D-GS) has emerged as an efficient 3D representation and a promising foundation for semantic tasks like segmentation. However, existing 3D-GS-based segmentation methods typically rely on high-dimensional category features, which introduce substantial memory overhead. Moreover, fine-grained segmentation remains challenging due to label space congestion and the lack of stable multi-granularity control mechanisms. To address these limitations, we propose a coarse-to-fine binary encoding scheme for per-Gaussian category representation, which compresses each feature into a single integer via the binary-to-decimal mapping, drastically reducing memory usage. We further design a progressive training strategy that decomposes panoptic segmentation into a series of independent sub-tasks, reducing inter-class conflicts and thereby enhancing fine-grained segmentation capability. Additionally, we fine-tune opacity during segmentation training to address the incompatibility between photometric rendering and semantic segmentation, which often leads to foreground-background confusion. Extensive experiments on multiple benchmarks demonstrate that our method achieves state-of-the-art segmentation performance while significantly reducing memory consumption and accelerating inference.

We propose a modular speaker diarization framework that collaborates with front-end techniques in a three-stage process, designed for the challenging CHiME-8 NOTSOFAR-1 acoustic environment. The framework leverages the strengths of deep learning based speech separation systems and traditional speech signal processing techniques to provide more accurate initializations for the Neural Speaker Diarization (NSD) system at each stage, thereby enhancing the performance of a single-channel NSD system. Firstly, speaker overlap detection and Continuous Speech Separation (CSS) are applied to the multichannel speech to obtain clearer single-speaker speech segments for the Clustering-based Speaker Diarization (CSD), followed by the first NSD decoding. Next, the binary speaker masks from the first decoding are used to initialize a complex Angular Center Gaussian Mixture Model (cACGMM) to estimate speaker masks on the multi-channel speech. Using Mask-to-VAD post-processing techniques, we achieve per-speaker speech activity with reduced speaker error (SpkErr), followed by a second NSD decoding. Finally, the second decoding results are used to Guide Source Separation (GSS) to produce per-speaker speech segments. Short utterances containing one word or fewer are filtered, and the remaining speech segments are re-clustered for the final NSD decoding. We present evaluation results progressively explored from the CHiME-8 NOTSOFAR-1 challenge, demonstrating the effectiveness of our modular diarization system and its contribution to improving speech recognition performance. The code will be open-sourced at https://github.com/rywang99/USTC-NERCSLIP_CHiME-8 . • We propose a novel three-stage modular speaker diarization framework integrating front-end cues. • CSS streams extend single-speaker segments for CSD clustering and NSD decoding initialization. • Spatial information is utilized to progressively reduce SpkErr and improve ASR performance.

The primary objective of Optical Chemical Structure Recognition is to identify chemical structure images into corresponding markup sequences. However, the complex two-dimensional structures of molecules, particularly those with rings and multiple branches, present significant challenges for current end-to-end methods to learn one-dimensional markup directly. To overcome this limitation, we propose a novel Ring-Free Language (RFL), which utilizes a divide-and-conquer strategy to describe chemical structures in a hierarchical form. RFL allows complex molecular structures to be decomposed into multiple parts, ensuring both uniqueness and conciseness while enhancing readability. This approach significantly reduces the learning difficulty for recognition models. Leveraging RFL, we propose a universal Molecular Skeleton Decoder (MSD), which comprises a skeleton generation module that progressively predicts the molecular skeleton and individual rings, along with a branch classification module for predicting branch information. Experimental results demonstrate that the proposed RFL and MSD can be applied to various mainstream methods, achieving superior performance compared to state-of-the-art approaches in both printed and handwritten scenarios.

In recent years, visually-rich document understanding has attracted increasing attention. Transformer-based pre-trained models have become the mainstream approach, yielding significant performance gains in this field. However, the self-attention mechanism's quadratic computational complexity hinders their efficiency and ability to process long documents. In this paper, we present DocMamba, a novel framework based on the state space model. It is designed to reduce computational complexity to linear while preserving global modeling capabilities. To further enhance its effectiveness in document processing, we introduce the Segment-First Bidirectional Scan (SFBS) to capture contiguous semantic information. Experimental results demonstrate that DocMamba achieves new state-of-the-art results on downstream datasets such as FUNSD, CORD, and SORIE, while significantly improving speed and reducing memory usage. Notably, experiments on the HRDoc confirm DocMamba's potential for length extrapolation.

HangChen HangChen, Chao-Han Huck Yang, Jia-Chen Gu, Sabato Marco Siniscalchi, Jun Du. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

This paper presents a Multi-Modal EnvironmentAware Network (MEAN-RIR), which uses an encoder-decoder framework to predict room impulse response (RIR) based on multi-level environmental information from audio, visual, and textual sources. Specifically, reverberant speech capturing room acoustic properties serves as the primary input, which is combined with panoramic images and text descriptions as supplementary inputs. Each input is processed by its respective encoder, and the outputs are fed into cross-attention modules to enable effective interaction between different modalities. The MEAN-RIR decoder generates two distinct components: the first component captures the direct sound and early reflections, while the second produces masks that modulate learnable filtered noise to synthesize the late reverberation. These two components are mixed to reconstruct the final RIR. The results show that MEANRIR significantly improves RIR estimation, with notable gains in acoustic parameters.

Diffusion models have revolutionized the field of talking head generation, yet still face challenges in expressiveness, controllability, and stability in long-time generation. In this research, we propose an EmotiveTalk framework to address these issues. Firstly, to realize better control over the generation of lip movement and facial expression, a Vision-guided Audio Information Decoupling (V-AID) approach is designed to generate audio-based decoupled representations aligned with lip movements and expression. Specifically, to achieve alignment between audio and facial expression representation spaces, we present a Diffusion-based Co-speech Temporal Expansion (Di-CTE) module within V-AID to generate expression-related representations under multi-source emotion condition constraints. Then we propose a well-designed Emotional Talking Head Diffusion (ETHD) backbone to efficiently generate highly expressive talking head videos, which contains an Expression Decoupling Injection (EDI) module to automatically decouple the expressions from reference portraits while integrating the target expression information, achieving more expressive generation performance. Experimental results show that EmotiveTalk can generate expressive talking head videos, ensuring the promised controllability of emotions and metric stability during long-time generation, yielding state-of-the-art performance compared to existing methods. The main page of our paper can be found in https://emotivetalk.github.io/.

Deep neural network (DNN)-based lipreading models have achieved excellent recognition accuracy but are currently facing challenges related to user privacy. To address this, we propose a novel hybrid quantum-classical neural network (HQCNN) for lipreading that balances superior performance with enhanced privacy protection. The HQCNN-based lipreading model features an innovative variational quantum circuit (VQC) back-end, which transforms the output of the DNN front-end into quantum representations and predicts the posterior probability of each word. Furthermore, we introduce projection-valued encoding (PVE) and projection-valued measurement (PVM), enabling the VQC to handle inputs and outputs of dimensions that scale exponentially with the number of qubits, thereby substantially increasing its expressive power. Additionally, we explore the privacy-preserving properties of the HQCNN-based lipreading model by integrating differentially private stochastic gradient descent (DP-SGD). Experiments conducted on the LRW dataset demonstrate the model’s exceptional recognition accuracy and privacy-preserving capabilities.

User-defined keyword spotting (KWS) enhances the user experience by allowing individuals to customize keywords. However, in open-vocabulary scenarios, most existing methods commonly suffer from high false alarm rates with confusable words and are limited to either audio-only or text-only enrollment. Therefore, in this paper, we first explore the model’s robustness against confusable words. Specifically, we propose Phoneme-Level Contrastive Learning (PLCL), which refines and aligns query and source feature representations at the phoneme level. This method enhances the model’s disambiguation capability through fine-grained positive and negative comparisons for more accurate alignment, and it is generalizable to jointly optimize both audio-text and audio-audio matching, adapting to various enrollment modes. Furthermore, we maintain a context-agnostic phoneme memory bank to construct confusable negatives for data augmentation. Based on this, a third-category discriminator is specifically designed to distinguish hard negatives. Overall, we develop a robust and flexible KWS system, supporting different modality enrollment methods within a unified framework. Verified on the LibriPhrase dataset, the proposed approach achieves state-of-the-art performance.

Online Handwritten Text Recognition (OLHTR) has gained considerable attention for its diverse range of applications. Current approaches usually treat OLHTR as a sequence recognition task, employing either a single trajectory or image encoder, or multi-stream encoders, combined with a CTC or attention-based recognition decoder. However, these approaches face several drawbacks: 1) single encoders typically focus on either local trajectories or visual regions, lacking the ability to dynamically capture relevant global features in challenging cases; 2) multi-stream encoders, while more comprehensive, suffer from complex structures and increased inference costs. To tackle this, we propose a Collaborative learning-based OLHTR framework, called Col-OLHTR, that learns multimodal features during training while maintaining a single-stream inference process. Col-OLHTR consists of a trajectory encoder, a Point-to-Spatial Alignment (P2SA) module, and an attention-based decoder. The P2SA module is designed to learn image-level spatial features through trajectory-encoded features and 2D rotary position embeddings. During training, an additional image-stream encoder-decoder is collaboratively trained to provide supervision for P2SA features. At inference, the extra streams are discarded, and only the P2SA module is used and merged before the decoder, simplifying the process while preserving high performance. Extensive experimental results on several OLHTR benchmarks demonstrate the state-of-the-art (SOTA) performance, proving the effectiveness and robustness of our design.

Although fully end-to-end speaker diarization systems have made significant progress in recent years, modular systems often achieve superior results in real-world scenarios due to their greater adaptability and robustness. Historically, modular speaker diarization methods have seldom discussed how to leverage spatial cues from multi-channel speech. This paper proposes a three-stage modular system to enhance single-channel neural speaker diarization systems and recognition performance by utilizing spatial cues from multi-channel speech to provide more accurate initialization for each stage of neural speaker diarization (NSD) decoding: (1) Overlap detection and continuous speech separation (CSS) on multi-channel speech are used to obtain cleaner single speaker speech segments for clustering, followed by the first NSD decoding pass. (2) The results from the first pass initialize a complex Angular Central Gaussian Mixture Model (cACGMM) to estimate speaker-wise masks on multi-channel speech, and through Overlap-add and Mask-to-VAD, achieve initialization with lower speaker error (SpkErr), followed by the second NSD decoding pass. (3) The second decoding results are used for guided source separation (GSS), recognizing and filtering short segments containing less one word to obtain cleaner speech segments, followed by re-clustering and the final NSD decoding pass. We presented the progressively explored evaluation results from the CHiME-8 NOTSOFAR-1 (Natural Office Talkers in Settings Of Far-field Audio Recordings) challenge, demonstrating the effectiveness of our system and its contribution to improving recognition performance. Our final system achieved the first place in the challenge.

In multimodal sentiment analysis, collecting text data is often more challenging than video or audio due to higher annotation costs and inconsistent automatic speech recognition (ASR) quality. To address this challenge, our study has developed a robust model that effectively integrates multimodal sentiment information, even in the absence of text modality. Specifically, we have developed a Double-Flow Self-Distillation Framework, including Unified Modality Cross-Attention (UMCA) and Modality Imagination Autoencoder (MIA), which excels at processing both scenarios with complete modalities and those with missing text modality. In detail, when the text modality is missing, our framework uses the LLM-based model to simulate the text representation from the audio modality, while the MIA module supplements information from the other two modalities to make the simulated text representation similar to the real text representation. To further align the simulated and real representations, and to enable the model to capture the continuous nature of sample orders in sentiment valence regression tasks, we have also introduced the Rank-N Contrast (RNC) loss function. When testing on the CMU-MOSEI, our model achieved outstanding performance on MAE and significantly outperformed other models when text modality is missing. The code is available at: https://github.com/WarmCongee/SDUMC.

Talking head generation intends to produce vivid and realistic talking head videos from a single portrait and speech audio clip. Although significant progress has been made in diffusion-based talking head generation, almost all methods rely on autoregressive strategies, which suffer from limited context utilization beyond the current generation step, error accumulation, and slower generation speed. To address these challenges, we present DAWN (\textbf{D}ynamic frame \textbf{A}vatar \textbf{W}ith \textbf{N}on-autoregressive diffusion), a framework that enables all-at-once generation of dynamic-length video sequences. Specifically, it consists of two main components: (1) audio-driven holistic facial dynamics generation in the latent motion space, and (2) audio-driven head pose and blink generation. Extensive experiments demonstrate that our method generates authentic and vivid videos with precise lip motions, and natural pose/blink movements. Additionally, with a high generation speed, DAWN possesses strong extrapolation capabilities, ensuring the stable production of high-quality long videos. These results highlight the considerable promise and potential impact of DAWN in the field of talking head video generation. Furthermore, we hope that DAWN sparks further exploration of non-autoregressive approaches in diffusion models. Our code will be publicly available at \url{https://github.com/Hanbo-Cheng/DAWN-pytorch}.

Sound event localization and detection (SELD) is a task that involves identifying and locating sound events in a given environment, which combines sound event detection (SED) and direction-of-arrival (DOA) estimation. This study addresses the extended task of audio-visual (AV) sound event localization and detection with source distance estimation (3D SELD). To leverage effective visual information, we propose an audio-guided visual attention mechanism to extract location-based features. We use two methods to fuse audio and visual features. Additionally, we introduce a source coordinate estimation (SCE) task that integrates DOA and distance estimation. Experimental results demonstrate that our proposed model significantly outperforms the official audio-only and AV baselines of the DCASE 2024 Challenge Task 3 on the development set of the STARSS23 dataset, even surpassing the challenge’s winning method. Attention visualization further highlights the effectiveness of audio information in localizing sound sources within visual images, ultimately enhancing the 3D SELD performance. Codes are available at https://github.com/qingwang24/AGVA-3DSELD/.

Existing text-to-music generation models face two main limitations: they either only generate approximately 10 seconds of music, falling short of users’ needs for longer compositions, or they require longer-duration datasets and increased output feature dimensions for training, leading to higher data and computational costs. We systematically explored two training-free strategies for long-duration music generation based on diffusion models. Both methods divide the task into multiple stages, where each stage generates a segment of music. The outputs from each stage are then concatenated to form the complete composition. The first, called the audio-prompt-guided strategy, employs the tail of audio features from the previous stage to serve as a prefix of the current stage, using masking techniques to guide the generation of the remainder of the current stage. The second, the structure-guided strategy, leverages the structure of previous audio representations as intermediate states in the current stage generation, initiating the inference sampling process from the intermediate time step. Our experiments demonstrate that existing short-form music generation systems are capable of generating long-duration music. Our demo page is provided at https://longmusic.github.io/long-music/.

Text-to-music (TTM) generation, which converts textual descriptions into audio, opens up innovative avenues for multimedia creation. Achieving high quality and diversity in this process demands extensive, high-quality data, which are often scarce in available datasets. Most open-source datasets frequently suffer from issues like low-quality waveforms and low text-audio consistency, hindering the advancement of music generation models. To address these challenges, we propose a novel quality-aware training paradigm for generating high-quality, high-musicality music from large-scale, quality-imbalanced datasets. Additionally, by leveraging unique properties in the latent space of musical signals, we adapt and implement a masked diffusion transformer (MDT) model for the TTM task, showcasing its capacity for quality control and enhanced musicality. Furthermore, we introduce a three-stage caption refinement approach to address low-quality captions' issue. Experiments show state-of-the-art (SOTA) performance on benchmark datasets including MusicCaps and the Song-Describer Dataset with both objective and subjective metrics. Demo audio samples are available at https://qa-mdt.github.io/, code and pretrained checkpoints are open-sourced at https://github.com/ivcylc/OpenMusic.

Recent advances in Multimodal Large Language Models (MLLMs) have achieved remarkable progress in general domains and demonstrated promise in multimodal mathematical reasoning. However, applying MLLMs to geometry problem solving (GPS) remains challenging due to lack of accurate step-by-step solution data and severe hallucinations during reasoning. In this paper, we propose GeoGen, a pipeline that can automatically generates step-wise reasoning paths for geometry diagrams. By leveraging the precise symbolic reasoning, GeoGen produces large-scale, high-quality question-answer pairs. To further enhance the logical reasoning ability of MLLMs, we train GeoLogic, a Large Language Model (LLM) using synthetic data generated by GeoGen. Serving as a bridge between natural language and symbolic systems, GeoLogic enables symbolic tools to help verifying MLLM outputs, making the reasoning process more rigorous and alleviating hallucinations. Experimental results show that our approach consistently improves the performance of MLLMs, achieving remarkable results on benchmarks for geometric reasoning tasks. This improvement stems from our integration of the strengths of LLMs and symbolic systems, which enables a more reliable and interpretable approach for the GPS task. Codes are available at https://github.com/ycpNotFound/GeoGen.

Recent rapid progress in Text-to-Audio (T2A) models contrasts sharply with the stagnation observed in the evolution of corresponding evaluation benchmarks. Existing benchmarks, such as AudioCaps, suffer from limited diversity and quality, as well as biased category distributions, leading to increasingly questionable reliability in assessing advanced T2A models. This paper introduces AudioAtlas, a comprehensive and balanced evaluation benchmark specifically designed for evaluating T2A models aimed at movie production. Based on an object-centric audio category system, AudioAtlas provides high-quality reference samples characterized by categorical balance and diversity. It includes detailed overall and event-level captions with rich descriptors, plus fine-grained temporal annotations from human experts, enabling thorough evaluation of temporal alignment and semantic accuracy. To enable precise evaluation of temporally-aligned generation across universal categories, two novel metrics are proposed leveraging recent advancements in large-scale Audio Language Models (AudioLLMs) and contrastive learning models. By re-benchmarking six currently influential T2A models, AudioAtlas provides evaluations better aligned with aesthetic considerations, offering clearer optimization directions for movie-production-oriented T2A systems. Additionally, we conduct a comprehensive comparative analysis on temporally-controllable T2A methods with training-based, and promising training-free approaches inspired by region-controllable image generation, clarifying current limitations and pointing out directions for future research. Audio specifically refers to sound event excluding speech and music. Further details are available on the project page: https://audioatlas.github.io/AudioAtlas/

We propose a novel approach to speech enhancement, termed Controllable ConforMer for Speech Enhancement (CCMSE), which leverages a Conformer-based architecture integrated with a control factor embedding module. Our method is designed to optimize speech quality for both human auditory perception and automatic speech recognition (ASR). It is observed that while mild denoising typically preserves speech naturalness, stronger denoising can improve human auditory tasks but often at the cost of ASR accuracy due to increased distortion. To address this, we introduce an algorithm that balances these trade-offs. By utilizing differential equations to interpolate between outputs at varying levels of denoising intensity, our method effectively combines the robustness of mild denoising with the clarity of stronger denoising, resulting in enhanced speech that is well-suited for both human and machine listeners. Experimental results on the CHiME-4 dataset validate the effectiveness of our approach.

Audio codecs, by discretizing continuous audio signals into finite token sets, achieve high-quality reconstruction at low bitrates in clean environments. However, real-world speech often deviates from ideal conditions, particularly in noisy environments with low signal-to-noise ratios (SNRs), limiting the performance of existing codecs in restoring clean audio from noisy inputs. To address this challenge, this paper introduces a Dual-Branch Codec (DB-Codec). Leveraging the hierarchical decomposition capability of residual vector quantization (RVQ), we separate noise and speech into codebooks at different layers through dual-branch reconstruction and orthogonality constraints between noise and speech features. DB-Codec integrates enhancement and synthesis into a unified model, enabling flexible control over noise suppression or signal recovery at equivalent compression rates to conventional codecs. Experiments demonstrate that our DB-Codec achieve an average improvement of 0.83 in PESQ and 9.16 in STOI compared to traditional codecs under low SNR conditions.

Audio-Visual Wake Word Spotting (AVWWS) aims to accurately detect user-defined keywords by leveraging the complementary nature of different modalities in challenging acoustic environments. However, two primary challenges hinder the application of AVWWS models in real-world scenarios: increased model parameters involving the video modality and the scarcity of paired audio-visual data. To address these issues, we propose a novel diverse acoustic knowledge distillation (DAKD) framework, which utilizes easily accessible single-modality audio data to train two teacher models and employs cross-modal knowledge distillation to transfer the generalization and de-noising capabilities of the teachers to the audio-visual student model. This approach mitigates the overfitting risk associated with large parameter counts and limited data. The DAKD framework consists of an audio-visual student model based on the lightweight multi-scale temporal-spatial attention (LMTSA) architecture, a multi-conditional teacher (MCT) model, and a de-noising teacher (DNT) model. The LMTSA model integrates compact 3D and 2D blocks based on the ResNet architecture through a simple attention module and accepts multi-scale supervision from word-level and phone-level labels, achieving joint temporal-spatial modeling with minimal parameter usage. The MCT and DNT models were trained using extensive real or simulated far-field speech and paired near-field and far-field speech, respectively, to generalize unseen acoustic environments and de-noising capabilities to the audio-visual student model. The effectiveness of our proposed DAKD framework is validated through comprehensive experiments on the MISP2021 and the updated MISP2021 Eval Hard datasets, establishing new benchmarks with fewer parameters. Our code will be available athttps://github.com/wikkk-tp/AVWWS_DAKD.

In this paper, we propose a novel approach for segmenting and tokenizing a video scene recording into a sequence of cascade units, known as visual segment units and modeled with visual segment models (VSMs) for video scene classification (VSC). Specifically, the proposed VSM framework takes deep visual features extracted from pre-trained encoders as inputs and models the temporal interactions between segment units by hidden Markov models. Next, we use unit co-occurrence statistics to introduce relationships between VSM units within a video scene recording. Furthermore, the VSM approach is extended to an acoustic-visual variant, subsequently integrating itself into a deep learning-based multi-modal scene classification system. This combination serves to further exploit the complementary nature of audio and video data. By incorporating a set of visual segment units into modeling a video scene class, it captures both inter-class similarity and intra-class diversity, facilitating improved scene classification, especially within categories prone to confusion. Extensive experimental results on a benchmark published by the DCASE (Detection and Classification of Acoustic Scenes and Events) 2021 Challenge show that the proposed framework can effectively handle the confusion issue among similar video scenes. In addition, our multi-modal integration system achieves state-of-the-art performance in the audio-visual scene classification task in the DCASE 2021 Challenge, thereby demonstrating the effectiveness of our proposed approach.

Advanced Audio- Visual Speech Recognition (AVSR) sys-tems have been observed to be sensitive to missing video frames, performing even worse than single-modality mod-els. While applying the common dropout techniques to the video modality enhances robustness to missing frames, it simultaneously results in a performance loss when dealing with complete data input. In this study, we delve into this contrasting phenomenon through the lens of modality bias and uncover that an excessive modality bias towards the audio modality induced by dropout constitutes the fun-damental cause. Next, we present the Modality Bias Hy-pothesis (MBH) to systematically describe the relationship between the modality bias and the robustness against missing modality in multimodal systems. Building on these findings, we propose a novel Multimodal Distribution Approxi-mation with Knowledge Distillation (MDA-KD)framework to reduce over-reliance on the audio modality, maintaining performance and robustness simultaneously. Finally, to address an entirely missing modality, we adopt adapters to dynamically switch decision strategies. The effective-ness of our proposed approach is evaluated through comprehensive experiments on the MISP2021 and MISP2022 datasets. Our code is available at https://github.com/dalision/ModalBiasAV5R.

Handwritten Mathematical Expression Recognition (HMER) forms a crucial task in the domain of document intelligence. It encompasses online and offline modalities, which utilize the trajectory sequence and static image as input, respectively. It is intuitive to utilize both online and offline modalities to build a more powerful recognition system. However, a formidable challenge arises as a result of the substantial heterogeneity between the online and offline modalities, which consequently leads to considerable obstacles in their alignment and fusion. In this work, we perceive the writing process as a video and introduce the Aggregated Optical Flow Map (AOFM) to represent the online modality, which is more compatible with the offline modality. Additionally, we propose the Optical Flow Aware Network (OFAN) in order to automatically extract, align, and fuse the features across online and offline modalities. Through experiment analysis, our method can be seamlessly applied to multiple existing offline HMER models, thereby yielding stable and substantial enhancements across CROHME 2014, 2016, and 2019 datasets. The code in this work is available at https: //github.com/Hanbo-Cheng/OFAN.git.

Deep learning (DL)-based speaker diarization methods have proven powerful performance comparing to traditional clustering-based methods for multi-talker speech diarization and recognition in farfield scenes. However, most DL-based approaches cannot utilize the spatial information well due to the poor robustness to unknown array topology and acoustic scenario. In this paper, a spatial long-term iterative mask estimation (SLT-IME) method is proposed to improve the performance of speaker diarization in various real-world acoustic scenarios. First, the complex angular central gaussian mixture model (cACGMM) with diarization results as initial values is used to estimate the presence probability of each speaker at each time-frequency bin, namely speaker masks, in a long-term chunk. Then, the speaker masks are converted to speaker activities according to the threshold, which deliver the diarization information of which speaker is active and when. Finally, the estimated speaker activity can also serve as the initial input for the diarization system, resulting in improved ASR performance. Experimental results on the CHiME-7 three datasets (CHiME-6, DiPCo, Mixer 6) show proposed method can improve diarization and recognition systems performance simultaneously. It also plays a key role in the ensemble system that achieves the best performance in the main track of CHiME-7 DASR Challenge.

In this study, we aim to improve our recent hierarchical information fusion system for multi-modal emotion recognition challenge (MER 2023) in both efficiency and performance. Specifically, we extract robust acoustic and visual representations from pre-trained models and fuse them together in different structures. Then, an entropy-based fusion approach is proposed to obtain the final prediction of emotion and valence based on multi-label predictions of all different feature fusion structures. Furthermore, to reduce the network redundancy and improve the model generalization in low-resource multi-modal data conditions, we propose a novel approach for optimizing the network structure progressively based on structured pruning and learning-rate rewinding. When tested on the dataset of MER 2023, the optimized network structure with entropy-based fusion yields consistent and significant improvements, outperforming the champion system of the MER-MULTI sub-challenge.

In multi-speaker scenarios, automatic speech recognition (ASR) models rely on pre-processed audio after speaker separation. However, when the target speaker is not accurately separated, ASR models face limitations in reaching their peak performance. To address this issue, we propose a speaker-adaptive ASR framework that possesses more implicit target speaker enhancement capability by efficiently joint-optimized speaker recognition (SR) and ASR models. Our framework introduces sharing self-supervised learning representation, optimization transfer and hierarchy speaker-gated attention. In this manner, it can maximize effectiveness of embedding bias and emphasize target speaker corresponding to semantic units. In the CHiME-7 DASR sub-track, the proposed method achieves a 28.19% relative reduction in word error rate (WER) on the development sets when compared to the official baseline. Notably, this framework has also been employed in the champion system for the CHiME-7 DASR.

Previous Multimodal Information based Speech Processing (MISP) challenges mainly focused on audio-visual speech recognition (AVSR) with commendable success. However, the most advanced back-end recognition systems often hit performance limits due to the complex acoustic environments. This has prompted a shift in focus towards the Audio-Visual Target Speaker Extraction (AVTSE) task for the MISP 2023 challenge in ICASSP 2024 Signal Processing Grand Challenges. Unlike existing audio-visual speech enhancement challenges primarily focused on simulation data, the MISP 2023 challenge uniquely explores how front-end speech processing, combined with visual clues, impacts back-end tasks in real-world scenarios. This pioneering effort aims to set the first benchmark for the AVTSE task, offering fresh insights into enhancing the accuracy of back-end speech recognition systems through AVTSE in challenging and real acoustic environments. This paper delivers a thorough overview of the task setting, dataset, and baseline system of the MISP 2023 challenge. It also includes an in-depth analysis of the challenges participants may encounter. The experimental results highlight the demanding nature of this task, and we look forward to the innovative solutions participants will bring forward.

We propose a novel neural speaker diarization system using memory-aware multi-speaker embedding with sequence-to-sequence architecture (NSD-MS2S), which integrates the strengths of memory-aware multi-speaker embedding (MA-MSE) and sequence-to-sequence (Seq2Seq) architecture, leading to improvement in both efficiency and performance. Next, we further decrease the memory occupation of decoding by incorporating input features fusion and then employ a multi-head attention mechanism to capture features at different levels. NSD-MS2S achieved a macro diarization error rate (DER) of 15.9% on the CHiME-7 EVAL set, which signifies a relative improvement of 49% over the official baseline system, and is the key technique for us to achieve the best performance for the main track of CHiME-7 DASR Challenge. Additionally, we introduce a deep interactive module (DIM) in MA-MSE module to better retrieve a cleaner and more discriminative multi-speaker embedding, enabling the current model to outperform the system we used in the CHiME-7 DASR Challenge. Our code is available at https://github.com/liyunlongaaa/NSD-MS2S.

Multimodal mathematical reasoning has gained increasing attention in recent times. However, previous effective methods have not tried to reason in the form of natural language. In this paper, we introduce a model named MATHS (MultimodAl Transformer-based Human-readable Solver) for visual arithmetic and geometry problems in multimodal mathematical reasoning tasks. Drawing inspiration from Multimodal Large Language Models (MLLMs), our approach involves generating problem-solving processes expressed in natural language, in order to leverage the inherent reasoning capabilities embedded within language models. To address the challenge of precise calculations for language models, our work proposes a Math-Constrained Generation (MCG) method to impose hard constraints on generated outputs. Extensive experiments demonstrate our model excels in visual arithmetic task, and achieves results that are either better or comparable to existing methods in geometry problems. Code is available at https://github.com/ycpNotFound/MATHS.

This study presents an audio-visual information fusion approach to sound event localization and detection (SELD) in low-resource scenarios. We aim at utilizing audio and video modality information through cross-modal learning and multi-modal fusion. First, we propose a cross-modal teacher-student learning (TSL) framework to transfer information from an audio-only teacher model, trained on a rich collection of audio data with multiple data augmentation techniques, to an audiovisual student model trained with only a limited set of multimodal data. Next, we propose a two-stage audio-visual fusion strategy, consisting of an early feature fusion and a late video-guided decision fusion to exploit synergies between audio and video modalities. Finally, we introduce an innovative video pixel swapping (VPS) technique to extend an audio channel swapping (ACS) method to an audio-visual joint augmentation. Evaluation results on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 Challenge data set demonstrate significant improvements in SELD performances. Furthermore, our submission to the SELD task of the DCASE 2023 Challenge ranks first place by effectively integrating the proposed techniques into a model ensemble.

In the previous Detection and Classification of Acoustic Scenes and Events (DCASE) Challenge Task 2: Anomalous Sound Detection (ASD) for Machine Condition Monitoring, each machine has a variety of different section IDs, which are subsets of the machine type. Therefore, section ID classification is often used to learn the representation of machine sounds for ASD. However, in real scenarios, it is both time-consuming and laborious for each machine to record data with multiple different section IDs. As such, the Task 2 of DCASE 2023 Challenge only includes one section ID for each machine, with the attribute information reflecting the machine’s working status and environment for recording. To this end, machine sound representations for ASD can be learned through the proxy task of two-stage multi-attribute classification. Specifically, the sounds of all machines are first used to pre-train a general attribute classification model. This model is then fine-tuned to obtain an attribute classification model specific to each machine, with a classification head established for each attribute that affects the acoustic characteristics of the machine in a multi-task learning framework. At the same time, data augmentation is used to improve the generalization capability caused by the limited amount of data in actual scenarios. Our approach demonstrates commendable performance on the Task 2 of DCASE 2023 Challenge. We further illustrate the effectiveness of our method through visual analysis.

Accurately identifying and organizing textual content is crucial for the automation of document processing in the field of form understanding. Existing datasets, such as FUNSD and XFUND, support entity classification and relationship prediction tasks but are typically limited to local and entity-level annotations. This limitation overlooks the hierarchically structured representation of documents, constraining comprehensive understanding of complex forms. To address this issue, we present the SRFUND, a hierarchically structured multi-task form understanding benchmark. SRFUND provides refined annotations on top of the original FUNSD and XFUND datasets, encompassing five tasks: (1) word to text-line merging, (2) text-line to entity merging, (3) entity category classification, (4) item table localization, and (5) entity-based full-document hierarchical structure recovery. We meticulously supplemented the original dataset with missing annotations at various levels of granularity and added detailed annotations for multi-item table regions within the forms. Additionally, we introduce global hierarchical structure dependencies for entity relation prediction tasks, surpassing traditional local key-value associations. The SRFUND dataset includes eight languages including English, Chinese, Japanese, German, French, Spanish, Italian, and Portuguese, making it a powerful tool for cross-lingual form understanding. Extensive experimental results demonstrate that the SRFUND dataset presents new challenges and significant opportunities in handling diverse layouts and global hierarchical structures of forms, thus providing deep insights into the field of form understanding. The original dataset and implementations of baseline methods are available at https://sprateam-ustc.github.io/SRFUND.

In recent years, the rapid advancement and widespread adoption of speech technology have made smart home systems a common feature in many households. However, individuals with dysarthria face difficulties using these technologies due to inconsistent speech patterns. This paper summarizes the Low-Resource Dysarthria Wake-Up Word Spotting (LRDWWS) Challenge at SLT 2024, which aimed to develop effective voice wake-up systems for individuals with dysarthria. The challenge attracted 25 teams from 4 countries, with 7 teams submitting results and 5 providing detailed system descriptions. This paper presents an overview of the dataset, evaluation metrics, and key innovations from participating teams. Our findings highlight the potential of these systems to enhance the accessibility and usability of smart home technologies for individuals with dysarthria. The challenge results underscore the importance of developing specialized solutions to meet the unique needs of this user group.

The StutteringSpeech Challenge focuses on advancing speech technologies for people who stutter, specifically targeting Stuttering Event Detection (SED) and Automatic Speech Recognition (ASR) in Mandarin. The challenge comprises three tracks: (1) SED, which aims to develop systems for detection of stuttering events; (2) ASR, which focuses on creating robust systems for recognizing stuttered speech; and (3) Research track for innovative approaches utilizing the provided dataset. We utilizes an open-source Mandarin stuttering dataset AS-70, which has been split into new training and test sets for the challenge. This paper presents the dataset, details the challenge tracks, and analyzes the performance of the top systems, highlighting improvements in detection accuracy and reductions in recognition error rates. Our findings underscore the potential of specialized models and augmentation strategies in developing stuttered speech technologies.

A multi-level distortion measure (MLDM) is proposed as an objective to optimize deep neural network-based speech enhancement (SE) in both audio-only and audio-visual scenarios. The aim is to achieve simultaneous performance improvements in speech quality, intelligibility, and recognition error reductions. Moreover, a comprehensive correlation analysis shows that these three evaluation metrics exhibit high Pearson correlation coefficient (PCC) values with three commonly used optimization objectives: the mean squared error between the ideal ratio and estimated magnitude masks, scale-invariant signal-to-noise ratio, and cross-entropy-guided measure. To further improve the performance, we leverage the complementarities of the three objectives and propose another correlated multi-level distortion measure (C-MLDM) defined as a weighted combination of MLDM and an average correlation measure based on the three PCCs. Experimental results on the TCD-TIMIT corpus corrupted by additive noise demonstrate that MLDM outperforms systems optimized with each objective in both audio-visual and audio-only scenarios, offering improved performances in all three metrics: speech quality, intelligibility, and recognition performance. C-MLDM also consistently outperforms MLDM in all test cases. Finally, the generalizability of both MLDM and C-MLDM is confirmed through extensive testing across diverse datasets, SE model architectures, and linguistic conditions. The source codes are publicly available.1

In this paper, we propose a variance-preserving interpolation framework to improve diffusion models for single-channel speech enhancement (SE) and automatic speech recognition (ASR). This new variance-preserving interpolation diffusion model (VPIDM) approach requires only 25 iterative steps and obviates the need for a corrector, an essential element in the existing variance-exploding interpolation diffusion model (VEIDM). Two notable distinctions between VPIDM and VEIDM are the scaling function of the mean of state variables and the constraint imposed on the variance relative to the mean's scale. We conduct a systematic exploration of the theoretical mechanism underlying VPIDM, and develop insights regarding VPIDM's applications in SE and ASR using VPIDM as a frontend. Our proposed approach, evaluated on two distinct data sets, demonstrates VPIDM's superior performances over conventional discriminative SE algorithms. Furthermore, we assess the performance of the proposed model under varying signal-to-noise ratio (SNR) levels. The investigation reveals VPIDM's improved robustness in target noise elimination when compared to VEIDM. Furthermore, utilizing the mid-outputs of both VPIDM and VEIDM results in enhanced ASR accuracies, thereby highlighting the practical efficacy of our proposed approach. Code and audio examples are available onlinehttps://github.com/zelokuo/VPIDM.

We propose a viseme subword modeling (VSM) approach to improve the generalizability and interpretability capabilities of deep neural network based lip reading. A comprehensive analysis of preliminary experimental results reveals the complementary nature of the conventional end-to-end (E2E) and proposed VSM frameworks, especially concerning speaker head movements. To increase lip reading accuracy, we propose hybrid viseme subwords and end-to-end modeling (HVSEM), which exploits the strengths of both approaches through multitask learning. As an extension to HVSEM, we also propose collaborative viseme subword and end-to-end modeling (CVSEM), which further explores the synergy between the VSM and E2E frameworks by integrating a state-mapped temporal mask (SMTM) into joint modeling. Experimental evaluations using different model backbones on both the LRW and LRW-1000 datasets confirm the superior performance and generalizability of the proposed frameworks. Specifically, VSM outperforms the baseline E2E framework, while HVSEM outperforms VSM in a hybrid combination of VSM and E2E modeling. Building on HVSEM, CVSEM further achieves impressive accuracies on 90.75% and 58.89%, setting new benchmarks for both datasets.

The problem of document structure reconstruction refers to converting digital or scanned documents into corresponding semantic structures. Most existing works mainly focus on splitting the boundary of each element in a single document page, neglecting the reconstruction of semantic structure in multi-page documents. This paper introduces hierarchical reconstruction of document structures as a novel task suitable for NLP and CV fields. To better evaluate the system performance on the new task, we built a large-scale dataset named HRDoc, which consists of 2,500 multi-page documents with nearly 2 million semantic units. Every document in HRDoc has line-level annotations including categories and relations obtained from rule-based extractors and human annotators. Moreover, we proposed an encoder-decoder-based hierarchical document structure parsing system (DSPS) to tackle this problem. By adopting a multi-modal bidirectional encoder and a structure-aware GRU decoder with soft-mask operation, the DSPS model surpass the baseline method by a large margin. All scripts and datasets will be made publicly available at https://github.com/jfma-USTC/HRDoc.

Most neural speaker diarization systems rely on sufficient manual training data labels, which are hard to collect under real-world scenarios. This paper proposes a semi-supervised speaker diarization system to utilize large-scale multi-channel training data by generating pseudo-labels for unlabeled data. Furthermore, we introduce cross-channel attention into the Neural Speaker Diarization Using Memory-Aware Multi-Speaker Embedding (NSD-MA-MSE) to learn channel contextual information of speaker embeddings better. Experimental results on the CHiME-7 Mixer6 dataset which only contains partial speakers’ labels of the training set, show that our system achieved 57.01% relative DER reduction compared to the clustering-based model on the development set. We further conducted experiments on the CHiME- 6 dataset to simulate the scenario of missing partial training set labels. When using 80% and 50% labeled training data, our system performs comparably to the results obtained using 100% labeled data for training.

The Multimodal Information based Speech Processing (MISP) 2022 challenge aimed to enhance speech processing performance in harsh acoustic environments by leveraging additional modalities such as video or text. The challenge included two tracks: audio-visual speaker diarization (AVSD) and audio-visual diarization and recognition (AVDR). The training material was based on previous MISP 2021 recordings, but we have accurately synchronized audio and visual data. Additionally, a new evaluation set was provided. This paper gives an overview of the challenge setup, presents the results, and summarizes the effective techniques employed by the participants. We also analyze the current technical challenges and suggest directions for future research in AVSD and AVDR.

The audio-visual direction of arrival (DOA) estimation has demonstrated superior performance recently. In this paper, we present a novel audio-visual multi-speaker DOA estimation network, which for the first time incorporates multi-speaker lip features to adapt the complex overlapping and noisy scenarios. Firstly, we encode the multi-channel audio features, the reference angles and the lip Regions of Interest (RoIs) detected from the video respectively to acquire high-level representations. Then the multi-modal embeddings of audio, speaker angles and lips are fused by a tri-modal gated fusion module to balance their contributions to the output. The fused embedding is sent to the backend network to obtain the accurate DOA estimation with the combination of the predicted speaker angular vectors and the speaker activities. Experimental results show that our proposed approach can reduce the localization error by 73.48% compared to the previous work on the 2021 Multi-modal Information based Speech Processing (MISP) Challenge corpus. Meanwhile, the high accuracy and stability of localization results demonstrate the robustness of the proposed model in multi-speaker scenarios.

We study four data augmentation (DA) techniques and two model architectures on realistic data for sound event localization and detection (SELD). First, based on ResNet-Conformer (RC), we compare the four DA approaches on the realistic DCASE 2022 SELD test set which is often not easy to handle due to room reverberations and audio overlaps in spontaneous recordings. Experimental results show that, except for audio channel swapping (ACS), the other three data augmentation methods that work well on the simulated SELD data set are no longer effective due to mismatches between simulated and realistic conditions. Next, using ACS-based augmentation, the two improved ResNet-Conformer networks further enhance SELD performances in realistic conditions. By incorporating these two sets of techniques, our overall system ranked the first place in SELD task of the DCASE 2022 Challenge.

In this paper, we introduce two further dilations of the recently introduced dilated nested arrays (DNAs), which possess an equal virtual ULA part to that of the nested arrays but possess two dense physical ULAs with the critical spacing (2 × λ/2). With having the parent array number of uniform DOFs unchanged, in the first dilation, the first dense ULA can be rearranged Qftimes, such that all the sensor pairs with the critical inter-sensor spacing therein are completely disposed of in a designated Qth-order dilated nested array, for 2 ≤ Q ≤ Qf+1, while in the second dilation called the super dilated nested array (SDNA), the second dense ULA in Qth-order DNAs is also rearranged, so as to have fixed weights as those of a co-prime array. Numerical examples are included to demonstrate the superior performance of these arrays.

The development of quantum machine learning demonstrates its quantum advantages over traditional deep learning, which promises to discover new patterns on supervised classification datasets. This work proposes a classical-to-quantum transfer learning system based on the large-scale unsupervised pre-trained model to demonstrate the competitive performance of quantum transfer learning for synthetic speech detection. We use the pre-trained model WavLM-Large to extract feature maps from speech signals, obtain low-dimensional embedding vectors through classical network components, and then jointly fine-tune the pre-trained model and classical network components with a variational quantum circuit (VQC). We evaluate our system on the ASVspoof 2021 DF task, and the experiments using quantum circuit simulations show that quantum transfer learning can improve the performance of the classical transfer learning baseline on the task.

This study focuses on the design of a loss function for a deep neural network (DNN)-based model with two branches, which is used to solve sound event localization and detection (SELD) on low-resource realistic data. To this end, we employ a secondary network for audio classification, which provides global event information to the main network, enabling it to make robust SELD predictions. Furthermore, we suggest utilizing a momentum strategy for direction-of-arrival (DOA) estimation, taking advantage of the strong temporal consistency of sound events, thereby effectively reducing localization error. Lastly, we incorporate a regularization term into the loss function to alleviate the overfitting problem on the small dataset. We evaluate our proposed methods on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2022 Task 3 dataset, and the results demonstrate consistent improvements in SELD performance. In comparison to the baseline system, the proposed loss function yields significantly improved results for both localization and detection metrics on realistic data. Moreover, the proposed loss function demonstrates its ability to generalize across different network architectures, as evidenced by the consistent improvements achieved.

The Multi-modal Information based Speech Processing (MISP) challenge aims to extend the application of signal processing technology in specific scenarios by promoting the research into wake-up words, speaker diarization, speech recognition, and other technologies. The MISP2022 challenge has two tracks: 1) audio-visual speaker diarization (AVSD), aiming to solve "who spoken when" using both audio and visual data; 2) a novel audio-visual diarization and recognition (AVDR) task that focuses on addressing "who spoken what when" with audio-visual speaker diarization results. Both tracks focus on the Chinese language, and use far-field audio and video in real home-tv scenarios: 2-6 people communicating each other with TV noise in the background. This paper introduces the dataset, track settings, and baselines of the MISP2022 challenge. Our analyses of experiments and examples indicate the good performance of AVDR baseline system, and the potential difficulties in this challenge due to, e.g., the far-field video quality, the presence of TV noise in the background, and the indistinguishable speakers.

Video information has been widely introduced to speech enhancement as its contribution at low signal-to-noise ratios (SNRs). Conventional audio-visual speech enhancement networks take noisy speech and video as input and learn features of clean speech directly. To reduce the large SNR gap between the learning target and input noisy speech, we propose a novel mask-based audio-visual progressive learning speech enhancement (AVPL) framework with visual information reconstruction (VIR) to increase SNRs gradually. Each stage of AVPL takes a concatenation of pre-trained visual embedding and the previous representation as input and predicts a mask with the intermediate representation of the current stage. To extract more visual information and deal with the performance distortion, the AVPL-VIR model reconstructs the visual embedding as it is fed in for each stage. Experiment on the TCD-TIMIT dataset shows that the progressive learning method significantly outperforms direct learning for both audio-only and audio-visual models. Moreover, by reconstructing video information, the VIR module provides a more accurate and comprehensive representation of the data, which in turn improves the performance of both AVDL and AVPL.

In recent research, slight performance improvement is observed from automatic speech recognition systems to audio-visual speech recognition systems in end-to-end frameworks with low-quality videos. Unmatching convergence rates and specialized input representations between audio-visual modalities are considered to cause the problem. In this paper, we propose two novel techniques to improve audio-visual speech recognition (AVSR) under a pre-training and fine-tuning training framework. First, we explore the correlation between lip shapes and syllable-level subword units in Mandarin through a frame-level subword unit classification task with visual streams as input. The fine-grained subword labels guide the network to capture temporal relationships between lip shapes and result in an accurate alignment between video and audio streams. Next, we propose an audio-guided Cross-Modal Fusion Encoder (CMFE) to utilize main training parameters for multiple cross-modal attention layers to make full use of modality complementarity. Experiments on the MISP2021-AVSR data set show the effectiveness of the two proposed techniques. Together, using only a relatively small amount of training data, the final system achieves better performances than state-of-the-art systems with more complex front-ends and back-ends. The code is released at1.

We propose an effective frame-level embedding learning framework for few-shot bioacoustic event detection (FSBED). First, the duration of different animal calls varies greatly, so we innovatively propose a frame-level embedding learning scheme, which can obtain adaptive event receptive fields with more accurate frame-level units. Next, we develop a transfer learning-based approach to deal with the mismatch between training and testing data. Finally, we use the idea of semi-supervised learning to solve the problem of too little labeled data in few-shot learning. By incorporating these several sets of techniques, our overall system ranked first place in the FSBED task of Detection and Classification of Acoustic Scenes and Events (DCASE) Challenge 2022.

Satisfactory recognition performance has been achieved for simple and controllable printed molecular images. However, recognizing handwritten chemical structure images remains unresolved due to the inherent ambiguities in handwritten atoms and bonds, as well as the signifcant challenge of converting projected 2D molecular layouts into markup strings. Target to address these problems, this paper proposes an end-to-end framework for handwritten chemical structure images recognition, with novel structure-specific markup language (SSML) and random conditional guided decoder (RCGD). SSML alleviates ambiguity and complexity in Chemfig syntax by designing an innovative markup language to accurately depict molecular structures. Besides, we propose RCGD to address the issue of multiple path decoding of molecular structures, which is composed of conditional attention guidance, memory classification and path selection mechanisms. In order to fully confirm the effectiveness of the end-to-end method, a new database containing 50,000 handwritten chemical structure images (EDU-CHEMC) has been established. Experimental results demonstrate that compared to traditional SMILES sequences, our SSML can significantly reduces the semantic gap between chemical images and markup strings. It is worth noting that our method can also recognize invalid or non-existent organic molecular structures, making it highly applicable for tasks related to teaching evaluations in the fields of chemistry and biology education. The EDU-CHEMC will be released soon in https://github.com/iFLYTEK-CV/EDU-CHEMC.

In this paper, we propose a novel framework for recognizing both discrete and dimensional emotions. In our framework, deep features extracted from foundation models are used as robust acoustic and visual representations of raw video. Three different structures based on attention-guided feature gathering (AFG) are designed for deep feature fusion. Then, we introduce a joint decoding structure for emotion classification and valence regression in the decoding stage. A multi-task loss based on uncertainty is also designed to optimize the whole process. Finally, by combining three different structures on the posterior probability level, we obtain the final predictions of discrete and dimensional emotions. When tested on the dataset of multimodal emotion recognition challenge (MER 2023), the proposed framework yields consistent improvements in both emotion classification and valence regression. Our final system achieves state-of-the-art performance and ranks third on the leaderboard on MER-MULTI sub-challenge.

In this paper, we propose a neural speaker diarization (NSD) network architecture consisting of three key components. First, a memory-aware multi-speaker embedding (MA-MSE) mechanism is proposed to facilitate a dynamical refinement of speaker embedding to reduce a potential data mismatch between the speaker embedding extraction and the NSD network. Next, a speaker selection procedure is introduced to handle situations where the detected number of speakers is different from the assumed speaker size in the NSD network. Finally, an adaptive procedure is proposed to improve the required prior information for the nonoverlap speech segments in a given utterance during each iteration. We call our proposed framework adaptive neural speaker diarization with memory-aware multi-speaker embedding (ANSD-MA-MSE). Our method improves diarization performance in realistic operating scenarios, such as adverse acoustic environments, domain mismatches, and a varying, rather than fixed, number of speakers. Having been tested on both the AMI corpus and the DIHARD-III evaluation sets, our proposed approach consistently outperforms other state-of-the-art techniques in diarization error rates, including the results reported by the best single-model system in the DIHARD-III challenge. Our code is publicly available athttps://github.com/Maokui-He/NSD-MA-MSE.

We improve iterative separation-based speaker diarization (ISSD) with quality-aware dynamic masking (QDM). We call the proposed framework QDM-SSD. Compared with ISSD, QDM-SSD enhances the simulated data used for model adaptation through QDM to alleviate the influence of errors in speaker priors. In addition to data quality purification, QDM-SSD also makes the adaptation data sparse by automatically adjusting speaker overlap ratios according to data quality. Furthermore, using a sliding window over the adaptation data, clean regions in speech segments can be better localized. Experiments on the two-speaker conversational telephone speech (CTS) corpus show that the proposed QDM-SSD framework can reduce the diarization error rate (DER) by 18.56% relatively compared with ISSD. Moreover, QDM-SSD is shown to generalize to other two-speaker non-conversation telephone speech data sets where ISSD fails to work. Finally, we demonstrate that QDM-SSD can serve as a front-end to improve the performances of back-end automatic speech recognition.

In this paper, we propose a novel four-stage data augmentation approach to ResNet-Conformer based acoustic modeling for sound event localization and detection (SELD). First, we explore two spatial augmentation techniques, namely audio channel swapping (ACS) and multi-channel simulation (MCS), to deal with data sparsity in SELD. ACS and MDS focus on augmenting the limited training data with expanding direction of arrival (DOA) representations such that the acoustic models trained with the augmented data are robust to localization variations of acoustic sources. Next, time-domain mixing (TDM) and time-frequency masking (TFM) are also investigated to deal with overlapping sound events and data diversity. Finally, ACS, MCS, TDM and TFM are combined in a step-by-step manner to form an effective four-stage data augmentation scheme. Tested on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2020 data set, our proposed augmentation approach greatly improves the system performance, ranking our submitted system in the first place in the SELD task of the DCASE 2020 Challenge. Furthermore, we employ a ResNet-Conformer architecture to model both global and local context dependencies of an audio sequence and win the first place in the DCASE 2022 SELD evaluations.

Wireless acoustic sensor network (WASN) has shown a superiority over conventional microphone arrays in many aspects. There exists an important tradeoff between the performance and power consumption, as usually the sensors are power driven with a limited amount of battery resource. Given a prescribed performance bound, in literature sensor selection (SS) and rate allocation (RA) methods can be leveraged to optimize the energy efficiency. In this work, we propose a joint rate allocation and sensor selection (RASS) approach to simultaneously optimize the sensor subset and rate distribution, which is formulated by minimizing the total transmission power in terms of selection and bit-rate variables and constraining the residual noise power. It can be shown that under a set of linear constraints on beamforming, the linearly-constrained minimum variance (LCMV) beamformer is the optimal noise reduction filter. Based on this, the RASS reduces to a mixed semi-definite and bilinear programming problem, which is then solved using a two-step algorithm. As the selection and bit-rate unknowns are bilinear, we first consider to optimize their product, resulting in an upper bound of RASS. Then, we use McCormick envelopes to relax the bilinear constraint, resulting in a linear program. The final selection and bit-rate solutions are obtained by posterior randomized rounding. It can be shown that SS and RA are special cases of the proposed RASS. Numerical results using simulated WASNs validate the power efficiency of the proposed method as well as the robustness against dynamic factors.

Speech enhancement shows an important necessity in many audio applications, particularly in noisy environments, where the speech quality needs to be improved. In this work, we consider the single-channel noise reduction (NR) problem from the conventional signal processing perspective. As conventional single-channel NR filters suffer from a serious speech distortion (SD) problem, we propose an SD weighted single-channel Wiener filter (SDW-SWF) in the short-time Fourier transform domain, which is obtained by minimizing the mean-square error (MSE) of the clean speech plus a$\mu$-weighted residual noise variance. Based on the generalized eigenvalue decomposition (GEVD) and rank-$r$approximation of the speech correlation matrix, the SDW-SWF can be written as a linear combination of eigenpairs, from which some special cases reduce to existing single-channel NR filters. As such, the proposed SDW-SWF has two parameters (i.e.,$\mu$and$r$) to tradeoff the MSE and SD. Then we theoretically analyze the impacts of the tradeoff parameters on the NR performance in SD, residual noise variance and the output signal-to-noise ratio (SNR). In addition, it is shown that the STFT-domain SDW-SWF can be further extended to the time domain, where the derived theorems still hold. Numerical results from several perspectives validate the effectiveness of the proposed method.

Existing researches on handwritten Chinese characters are mainly based on recognition network designed to solve the complex structure and numerous amount characteristics of Chinese characters. In this paper, we investigate Chinese characters from the perspective of error correction, which is to diagnose a handwritten character to be right or wrong and provide a feedback on error analysis. For this handwritten Chinese character error correction task, we define a benchmark by unifying both the evaluation metrics and data splits for the first time. Then we design a diagnosis system that includes decomposition, judgement and correction stages. Specifically, a novel tree-structure analysis network (TAN) is proposed to model a Chinese character as a tree layout, which mainly consists of a CNN-based encoder and a tree-structure based decoder. Using the predicted tree layout for judgement, correction operation is performed for the wrongly written characters to do error analysis. The correction stage is composed of three steps: fetch the ideal character, correct the errors and locate the errors. Additionally, we propose a novel bucketing mining strategy to apply triplet loss at radical level to alleviate feature dispersion. Experiments on handwritten character dataset demonstrate that our proposed TAN shows great superiority on all three metrics comparing with other state-of-the-art recognition models. Through quantitative analysis, TAN is proved to capture more accurate spatial position information than regular encoder-decoder models, showing better generalization ability.

Document intelligence as a relatively new research topic supports many business applications. Its main task is to automatically read, understand, and analyze documents. However, due to the diversity of formats (invoices, reports, forms, etc.) and layouts in documents, it is difficult to make machines understand documents. In this paper, we present the GraphDoc, a multimodal graph attention-based model for various document understanding tasks. GraphDoc is pre-trained in a multimodal framework by utilizing text, layout, and image information simultaneously. In a document, a text block relies heavily on its surrounding contexts, accordingly we inject the graph structure into the attention mechanism to form a graph attention layer so that each input node can only attend to its neighborhoods. The input nodes of each graph attention layer are composed of textual, visual, and positional features from semantically meaningful regions in a document image. We do the multimodal feature fusion of each node by the gate fusion layer. The contextualization between each node is modeled by the graph attention layer. GraphDoc learns a generic representation from only 320k unlabeled documents via the Masked Sentence Modeling task. Extensive experimental results on the publicly available datasets show that GraphDoc achieves state-of-the-art performance, which demonstrates the effectiveness of our proposed method.

In recent years, tree decoders become more popular than LaTeX string decoders in the field of handwritten mathematical expression recognition (HMER) as they can capture the hierarchical tree structure of mathematical expressions. However previous tree decoders converted the tree structure labels into a fixed and ordered sequence, which could not make full use of the diversified expression of tree labels. In this study, we propose a novel tree decoder (TDv2) to fully utilize the tree structure labels. Compared with previous tree decoders, this new model does not require a fixed priority for different branches of a node during training and inference, which can effectively improve the model generalization capability. The input and output of the model make full use of the tree structure label, so that there is no need to find the parent node in the decoding process, which simplifies the decoding process and adds a prior information to help predict the node. We verified the effectiveness of each part of the model through comprehensive ablation experiments and attention visualization analysis. On the authoritative CROHME 14/16/19 datasets, our method achieves the state-of-the-art results.

In this paper we discuss the rational of the Multi-model Information based Speech Processing (MISP) Challenge, and provide a detailed description of the data recorded, the two evaluation tasks and the corresponding baselines, followed by a summary of submitted systems and evaluation results. The MISP Challenge aims at tack-ling speech processing tasks in different scenarios by introducing information about an additional modality (e.g., video, or text), which will hopefully lead to better environmental and speaker robustness in realistic applications. In the first MISP challenge, two bench-mark datasets recorded in a real-home TV room with two reproducible open-source baseline systems have been released to promote research in audio-visual wake word spotting (AVWWS) and audio-visual speech recognition (AVSR). To our knowledge, MISP is the first open evaluation challenge to tackle real-world issues of AVWWS and AVSR in the home TV scenario.

We propose two improvements to target-speaker voice activity detection (TS-VAD), the core component in our proposed speaker diarization system that was submitted to the 2022 Multi-Channel Multi-Party Meeting Transcription (M2MeT) challenge. These techniques are designed to handle multi-speaker conversations in real-world meeting scenarios with high speaker-overlap ratios and under heavy reverberant and noisy condition. First, for data preparation and augmentation in training TS-VAD models, speech data containing both real meetings and simulated indoor conversations are used. Second, in refining results obtained after TS-VAD based decoding, we perform a series of post-processing steps to improve the VAD results needed to reduce diarization error rates (DERs). Tested on the ALIMEETING corpus, the newly released Mandarin meeting dataset used in M2MeT, we demonstrate that our proposed system can decrease the DER by up to 66.55/60.59% relatively when compared with classical clustering based diarization on the Eval/Test set.

Single-channel speech enhancement for automatic speech recognition (ASR) has been widely studied. However, most speech enhancement methods conduct over suppression and introduce distortion, which limits performance gains or even deteriorates the back-end performance. The key to solving this problem is preserving the integrity of speech while suppressing the background noises. There-fore, we propose a time domain progressive learning (TDPL) approach for speech enhancement and ASR. TDPL model consists of encoder, progressive enhancer and decoder. Both SNR-increased intermediate target with less speech distortion and clean target with better listening quality/intelligibility are learned, which are provided for ASR pre-processing and speech communication, respectively. Additionally, we also present an SNR constriction loss that is fit for TDPL to further improve ASR performance. We evaluate the proposed methods on CHiME-4 real evaluation set. The results show that the TDPL method significantly outperforms time domain speech enhancement methods and frequency domain progressive learning methods in ASR task, and the intermediate output of TDPL achieves a 36.3% relative word error rate reduction with a powerful ASR back-end without retraining. Moreover, the estimated clean output achieves certain improvement on CHiME-4 simulation evaluation set in terms of PESQ and STOI measures.

In this paper, we propose an iterative separation-based speaker diarization (ISSD) approach to cope with the realistic data conditions. In the proposed ISSD, we iteratively generate adaptation data ac-cording to speaker priors and fine-tune the separation model, which leads to a gradual performance improvement. To further reduce some unavoidable speaker detection errors due to some undesirable prior errors using simple ISSD, we utilize speaker embedding information and propose two post-processing techniques, namely, speaker filtering and speaker recovery. We evaluate the diarization performance on the two-speaker conversational telephone speech (CTS) data set from DIHARD-III Challenge. When compared to state-of-the-art clustering-based speaker diarization (CSD) system, the proposed ISSD approach combined with the two post-processing schemes yields a 47.72 % and 46.97 % relative diarization error rate reduction on the development and evaluation sets, respectively. ISSD is also one key contributing factor to the best-performing system in DIHARD-III Challenge.

In this paper, we present a new sparse co-prime array design that achieves a higher number of degrees-of-freedom for direction-of-arrival (DOA) estimation. The proposed array design adopts a sequence of displacements applied in a systematic procedure to the sensors of one of the two constituent sub-arrays of the prototype co-prime array. Accordingly, the applied displacements completely eliminate the redundant virtual sensors between the two sub-arrays and at the same time considerably increase their number of uniform DOFs. Moreover, with no cross-sensor redundancies, the new design is of ideal first weight functions, and hence, it resolves sources more robustly than other co-prime array designs in the presence of mutual coupling. Simulation results demonstrate the superior performance of the proposed co-prime structure.

Audio-only based wake word spotting (WWS) is challenging under noisy conditions due to the environmental interference in signal transmission. In this paper, we investigate on designing a compact audio-visual WWS system by utilizing the visual information to alleviate the degradation. Specifically, in order to use visual information, we first encode the detected lips to fixed-size vectors with MobileNet and concatenate them with acoustic features followed by the fusion network for WWS. However, the audio-visual model based on neural network requires a large footprint and a high computational complexity. To meet the application requirements, we introduce a neural network pruning strategy via the lottery ticket hypothesis in an iterative fine-tuning manner (LTH-IF), to the single-modal and multi-modal models, respectively. Tested on our in-house corpus for audio-visual WWS in a home TV scene, the proposed audiovisual system achieves significant performance improvements over the single-modality (audio-only or video-only) system under different noisy conditions. Moreover, LTH-IF pruning can largely reduce the network parameters and computations with no degradation of WWS performance, leading to a potential product solution for the TV wake-up scenario.

Table of contents (ToC) extraction aims to extract headings of different levels in documents to better understand the outline of the contents, which can be widely used for document understanding and information retrieval. Existing works often use hand-crafted features and predefined rule-based functions to detect headings and resolve the hierarchical relationship between headings. Both the benchmark and research based on deep learning are still limited. Accordingly, in this paper, we first introduce a standard dataset, HierDoc, including image samples from 650 documents of scientific papers with their content labels. Then we propose a novel end-to-end model by using the multimodal tree decoder (MTD) for ToC as a benchmark for HierDoc. The MTD model is mainly composed of three parts, namely encoder, classifier, and decoder. The encoder fuses the multimodality features of vision, text, and layout information for each entity of the document. Then the classifier recognizes and selects the heading entities. Next, to parse the hierarchical relationship between the heading entities, a tree-structured decoder is designed. To evaluate the performance, both the metric of tree-edit-distance similarity (TEDS) and F1-Measure are adopted. Finally, our MTD approach achieves an average TEDS of 87.2% and an average F1-Measure of 88.1% on the test set of HierDoc. The code and dataset will be released at: https://github.com/Pengfei-Hu/MTD.

Self-supervised visual pre-training has recently emerged in scene text recognition (STR), which designs the pretext tasks and takes unlabeled data as input to obtain useful representations for STR. However, most current self-supervised methods do not pay special attention to the importance of sequence awareness. Accordingly, we propose a novel self-supervised STR method based on contrastive predictive coding (STR-CPC), which regards a text instance as a sequence from left to right and captures the visual sequence correlation. Considering the information overlap problem within the feature map induced by the deep convolutional neural network (CNN) encoder, we design a widthwise causal convolution during model pre-training and a progressive recovery training strategy (PRTS) during model fine-tuning to improve the STR performance. Experiments on scene text show that our STR-CPC method outperforms the existing self-supervised methods, which testifies the advantage of visual sequence correlation for STR. Additionally, STR-CPC observably boosts performance compared with supervised training when the amount of labeled data decreases.

In this paper, we present the updated Audio-Visual Speech Recognition (AVSR) corpus of MISP2021 challenge, a large-scale audio-visual Chinese conversational corpus consisting of 141h audio and video data collected by far/middle/near microphones and far/middle cameras in 34 real-home TV rooms. To our best knowledge, our corpus is the first distant multi-microphone conversational Chinese audio-visual corpus and the first large vocabulary continuous Chinese lip-reading dataset in the adverse home-tv scenario. Moreover, we make a deep analysis of the corpus and conduct a comprehensive ablation study of all audio and video data in the audio-only/video-only/audiovisual systems. Error analysis shows video modality supplement acoustic information degraded by noise to reduce deletion errors and provide discriminative information in overlapping speech to reduce substitution errors. Finally, we also design a set of experiments such as frontend, data augmentation and end-to-end models for providing the direction of potential future work. The corpus and the code are released to promote the research not only in speech area but also for the computer vision area and cross-disciplinary research.

In this paper, we describe and release publicly the audio-visual wake word spotting (WWS) database in the MISP2021 Challenge, which covers a range of scenarios of audio and video data collected by near-, mid-, and far-field microphone arrays, and cameras, to create a shared and publicly available database for WWS. The database and the code 2 are released, which will be a valuable addition to the community for promoting WWS research using multi-modality information in realistic and complex conditions. Moreover, we investigated the different data augmentation methods for single modalities on an end-to-end WWS network. A set of audio-visual fusion experiments and analysis were conducted to observe the assistance from visual information to acoustic information based on different audio and video field configurations. The results showed that the fusion system generally improves over the single-modality (audio- or video-only) system, especially under complex noisy conditions.

To improve device robustness, a highly desirable key feature of a competitive data-driven acoustic scene classification (ASC) system, a novel two-stage system based on fully convolutional neural networks (CNNs) is proposed. Our two-stage system leverages on an ad-hoc score combination based on two CNN classifiers: (i) the first CNN classifies acoustic inputs into one of three broad classes, and (ii) the second CNN classifies the same inputs into one of ten finergrained classes. Three different CNN architectures are explored to implement the two-stage classifiers, and a frequency sub-sampling scheme is investigated. Moreover, novel data augmentation schemes for ASC are also investigated. Evaluated on DCASE 2020 Task 1a, our results show that the proposed ASC system attains a state-of-the-art accuracy on the development set, where our best system, a two-stage fusion of CNN ensembles, delivers a 81.9% average accuracy among multi-device test data, and it obtains a significant improvement on unseen devices. Finally, neural saliency analysis with class activation mapping (CAM) gives new insights on the patterns learnt by our models.

In this paper, we propose a progressive learning-based adaptive noise and speech estimation (PL-ANSE) method for speech preprocessing in noisy speech recognition, leveraging upon a frame-level noise tracking capability of improved minima controlled recursive averaging (IMCRA) and an utterance-level deep progressive learning of nonlinear interactions between speech and noise. First, a bi-directional long short-term memory model is adopted at each network layer to learn progressive ratio masks (PRMs) as targets with progressively increasing signal-to-noise ratios. Then, the estimated PRMs at the utterance level are combined within a conventional speech enhancement algorithm at the frame level for speech enhancement. Finally, the enhanced speech based on multi-level information fusion is directly fed into a speech recognition system to improve the recognition performance. Experiments show that our proposed approach can achieve a relative word error rate (WER) reduction of 22.1% when compared to results attained with unprocessed noisy speech (from 23.84% to 18.57%) on the CHiME-4 single-channel real test data.

A new hierarchical convolutional neural network-based autoencoder architecture called SEHAE (Speech Enhancement Hierarchical AutoEncoder) is introduced, in which the latent representation is decomposed into several parts that correspond to different scales. The model consists of three functionally different components. First, a stack of encoders generates a set of latent vectors that contain information from an increasingly larger receptive field. Second, the decoders construct the clean speech in a stage-wise and additive fashion, starting from a learned initial vector. The third component, which we call funnel networks, is tasked with "knitting" together the outputs of the previous decoder and the encoder to compute latent vectors for the next decoder. Several options for initial vectors are explored. Experiments show that SEHAE achieves significant improvements for the considered speech quality and intelligibility measures, outperforming a denoising autoencoder and other step-wise models. Furthermore, its internal workings are investigated using the intermediate results from the decoders.

The recent criteria of a preferable linear sparse array for a robust direction of arrival (DOA) estimation are the closed-form expression for its sensor locations, the large central uniform linear array (ULA) segment in its resulting co-array and the fewer sensor pairs with small separations in its configuration. This paper aims to introduce a new proposed sparse array that takes into account all these considerations. The new sparse array configuration is proposed based on utilizing the two coordinates of a linear axis (TCLA) to situate its sub-arrays. Compared to the (super) nested array having the same sensor number, the TCLA array owns the same number of uniform degrees of freedom (DOFs) but possesses less mutual coupling effects. These properties are quantitatively covered, and numerical simulations are included to demonstrate the superior performance of the proposed array.

In this paper, we present AISHELL-4, a sizable real-recorded Mandarin speech dataset collected by 8-channel circular microphone array for speech processing in conference scenario. The dataset consists of 211 recorded meeting sessions, each containing 4 to 8 speakers, with a total length of 120 hours. This dataset aims to bridge the advanced research on multi-speaker processing and the practical application scenario in three aspects. With real recorded meetings, AISHELL-4 provides realistic acoustics and rich natural speech characteristics in conversation such as short pause, speech overlap, quick speaker turn, noise, etc. Meanwhile, accurate transcription and speaker voice activity are provided for each meeting in AISHELL-4. This allows the researchers to explore different aspects in meeting processing, ranging from individual tasks such as speech front-end processing, speech recognition and speaker diarization, to multi-modality modeling and joint optimization of relevant tasks. Given most open source dataset for multi-speaker tasks are in English, AISHELL-4 is the only Mandarin dataset for conversation speech, providing additional value for data diversity in speech community. We also release a PyTorch-based training and evaluation framework as baseline system to promote reproducible research in this field.

Target-speaker voice activity detection (TS-VAD) has recently shown promising results for speaker diarization on highly overlapped speech. However, the original model requires a fixed (and known) number of speakers, which limits its application to real conversations. In this paper, we extend TS-VAD to speaker diarization with unknown numbers of speakers. This is achieved by two steps: first, an initial diarization system is applied for speaker number estimation, followed by TS-VAD network output masking according to this estimate. We further investigate different diarization methods, including clustering-based and region proposal networks, for estimating the initial i-vectors. Since these systems have complementary strengths, we propose a fusion-based method to combine frame-level decisions from the systems for an improved initialization. We demonstrate through experiments on variants of the LibriCSS meeting corpus that our proposed approach can improve the DER by up to 50\% relative across varying numbers of speakers. This improvement also results in better downstream ASR performance approaching that using oracle segments.

DIHARD III was the third in a series of speaker diarization challenges intended to improve the robustness of diarization systems to variability in recording equipment, noise conditions, and conversational domain. Speaker diarization was evaluated under two speech activity conditions (diarization from a reference speech activity vs. diarization from scratch) and 11 diverse domains. The domains span a range of recording conditions and interaction types, including read audio-books, meeting speech, clinical interviews, web videos, and, for the first time, conversational telephone speech. A total of 30 organizations (forming 21teams) from industry and academia submitted 499 valid system outputs. The evaluation results indicate that speaker diarization has improved markedly since DIHARD I, particularly for two-party interactions, but that for many domains (e.g., web video) the problem remains far from solved.

This paper presents our main contributions of acoustic modeling for multi-array multi-talker speech recognition in the CHiME-6 Challenge, exploring different strategies for acoustic data augmentation and neural network architectures. First, enhanced data from our front-end network preprocessing and spectral augmentation are investigated to be effective for improving speech recognition performance. Second, several neural network architectures are explored by different combinations of deep residual network (ResNet), factorized time delay neural network (TDNNF) and residual bidirectional long short-term memory (RBiLSTM). Finally, multiple acoustic models can be combined via minimum Bayes risk fusion. Compared with the official baseline acoustic model, the proposed solution can achieve a relatively word error rate reduction of 19% for the best single ASR system on the evaluation data, which is also one of main contributions to our top system for the Track 1 tasks of the CHiME-6 Challenge.

Multi-speaker speech recognition of unsegmented recordings has diverse applications such as meeting transcription and automatic subtitle generation. With technical advances in systems dealing with speech separation, speaker diarization, and automatic speech recognition (ASR) in the last decade, it has become possible to build pipelines that achieve reasonable error rates on this task. In this paper, we propose an end-to-end modular system for the LibriCSS meeting data, which combines independently trained separation, diarization, and recognition components, in that order. We study the effect of different state-of-the-art methods at each stage of the pipeline, and report results using task-specific metrics like SDR and DER, as well as downstream WER. Experiments indicate that the problem of overlapping speech for diarization and ASR can be effectively mitigated with the presence of a well-trained separation module. Our best system achieves a speaker-attributed WER of 12.7%, which is close to that of a non-overlapping ASR.

A new cross-entropy-guided measure (CEGM) is proposed to indirectly assess accuracies of automatic speech recognition (ASR) of degraded speech with a speech enhancement front-end and without directly performing ASR experiments. The proposed CEGM is calculated in three steps, namely: (1) a low-level representations via feature extraction, (2) a high-level nonlinear mapping using an acoustic model, and (3) a final CEGM calculation between the high-level representations of clean and enhanced speech. Specifically, state posterior probabilities from outputs of conventional hybrid acoustic model of the target ASR system are adopted as the high-level representations and a cross-entropy criterion is used to calculate the CEGM. Due to CEGM's differentiability, it can also be used to replace the conventional minimum mean squared error (MMSE) criterion as an objective function for deep neural network (DNN)-based speech enhancement. Therefore, the front-end enhancement model can be optimized towards improving the accuracies of the back-end ASR system. Experiments on single-channel CHiME-4 Challenge show that CEGM yields consistently the highest correlations with word error rate (WER) which is often costly to calculate, and achieves the most accurate assessment of ASR performance when compared to the perceptual evaluation metrics commonly used for assessing speech enhancement performance. Furthermore, CEGM-optimized speech enhancement could effectively reduce the WER on the CHiME-4 real test set when compared to unprocessed noisy speech and enhanced speech obtained with MMSE-optimized enhancement for ASR systems with fixed multi-condition acoustic models in various deep architectures.

For multi-microphone speech enhancement, different microphones might have different contributions, assome are even marginal. This is more likely to happen in wireless acoustic sensor networks (WASNs), where somesensors might be distant. In this work, we therefore consider sensor selection for linearly-constrained beamformers. Theproposed sensor selection approach is formulated by minimizing the total output noise power and constraining thenumber of selected sensors. As the considered sensor selection problem requires the relative acoustic transfer function(RTF), the covariance whitening based RTF estimation or a direct-path RTF approximation is exploited. For a singletarget source, we can thus substitute the estimated RTF or the assumed RTF to the original problem formulation in orderto design a minimum variance distortionless response (MVDR) beamformer. Alternatively, we can integrate the two RTFsto design a linearly constrained minimum variance (LCMV) beamformer in order to alleviate the effects of RTFestimation/approximation errors. By leveraging the superiority of LCMV beamformers, the proposed approach can beapplied to the multi-source case. An evaluation using a simulated large-scale WASN demonstrates that the integration ofRTFs for the sensor selection based LCMV beamformer can be beneficial as opposed to relying on either of theindividual RTF steered sensor selection based MVDR beamformers. We conclude that the sensors that are close to thetarget source(s) and also some around the coherent interferers are more informative.

Multimodal emotion recognition is a challenging task in emotion computing as it is quite difficult to extract discriminative features to identify the subtle differences in human emotions with abstract concept and multiple expressions. Moreover, how to fully utilize both audio and visual information is still an open problem. In this paper, we propose a novel multimodal fusion attention network for audio-visual emotion recognition based on adaptive and multi-level factorized bilinear pooling (FBP). First, for the audio stream, a fully convolutional network (FCN) equipped with 1-D attention mechanism and local response normalization is designed for speech emotion recognition. Next, a global FBP (G-FBP) approach is presented to perform audio-visual information fusion by integrating self-attention based video stream with the proposed audio stream. To improve G-FBP, an adaptive strategy (AG-FBP) to dynamically calculate the fusion weight of two modalities is devised based on the emotion-related representation vectors from the attention mechanism of respective modalities. Finally, to fully utilize the local emotion information, adaptive and multi-level FBP (AM-FBP) is introduced by combining both global-trunk and intra-trunk data in one recording on top of AG-FBP. Tested on the IEMOCAP corpus for speech emotion recognition with only audio stream, the new FCN method outperforms the state-of-the-art results with an accuracy of 71.40%. Moreover, validated on the AFEW database of EmotiW2019 sub-challenge and the IEMOCAP corpus for audio-visual emotion recognition, the proposed AM-FBP approach achieves the best accuracy of 63.09% and 75.49% respectively on the test set.

Recently, recognition of online handwritten mathe- matical expression has been greatly improved by employing encoder-decoder based methods. Existing encoder-decoder models use string decoders to generate LaTeX strings for mathematical expression recognition. However, in this paper, we importantly argue that string representations might not be the most natural for mathematical expressions – mathematical expressions are inherently tree structures other than flat strings. For this purpose, we propose a novel sequential relation decoder (SRD) that aims to decode expressions into tree structures for online handwritten mathematical expression recognition. At each step of tree construction, a sub-tree structure composed of a relation node and two symbol nodes is computed based on previous sub-tree structures. This is the first work that builds a tree structure based decoder for encoder-decoder based mathematical expression recognition. Compared with string decoders, a decoder that better understands tree structures is crucial for mathematical expression recognition as it brings a more reasonable learning objective and improves overall generalization ability. We demonstrate how the proposed SRD outperforms state-of-the-art string decoders through a set of experiments on CROHME database, which is currently the largest benchmark for online handwritten mathematical expression recognition.

The spectral information of acoustic scenes is diverse and complex, which poses challenges for acoustic scene tasks. To improve the classification performance, a variety of convolutional neural networks (CNNs) are proposed to extract richer semantic information of scene utterances. However, the different regions of the features extracted from CNN-based encoder have different importance. In this paper, we propose a novel strategy for acoustic scene classification, namely high-resolution attention network with acoustic segment model (HRAN-ASM). In this approach, we utilize fully CNN to obtain high-level semantic information and then adopt two-stage attention strategy to select the relevant acoustic scene segments. Besides, the acoustic segment model (ASM) proposed in our recent work provides embedding vectors for this attention mechanism. The performance is evaluated on DCASE 2018 Task 1a, showing 70.5% good classification accuracy under single system and no data expansion, which is superior to CNN-based self-attention mechanism and highly competitive.

In this paper, we propose a novel adaptive technique that uses an attention-based gated scaling (AGS) scheme to improve deep feature learning for connectionist temporal classification (CTC) acoustic modeling. In AGS, the outputs of each hidden layer of the main network are scaled by an auxiliary gate matrix extracted from the lower layer by using an attention mechanism. Furthermore, the auxiliary AGS layer and the main network are jointly trained without requiring second-pass model training or additional speaker information, such as i-vector. On the Mandarin AISHELL-1 dataset, the proposed AGS yields a 7.94% character error rate (CER). To the best of our knowledge, the results obtained when training on the full AISHELL-1 training set, are the best published currently for the end-to-end systems.

This paper presents an improved deep embedding learning method based on a convolutional neural network (CNN) for text-independent speaker verification. Two improvements are proposed for x-vector embedding learning: (1) a multiscale convolution (MSCNN) is adopted in the frame-level layers to capture the complementary speaker information in different receptive fields; (2) a Baum-Welch statistics attention (BWSA) mechanism is applied in the pooling layer, which can integrate more useful long-term speaker characteristics in the temporal pooling layer. Experiments are carried out on the NIST SRE16 evaluation set. The results demonstrate the effectiveness of the MSCNN and show that the proposed BWSA can further improve the performance of the DNN embedding system.

The multi-objective learning using minimum mean squared error criterion for DNN-based speech enhancement (MMSE-MOL-DNN) has been demonstrated to achieve better performance than single output DNN. However, one problem of MMSE-MOL-DNN is that the prediction error values on different targets have a very broad dynamic range, causing difficulty in DNN training. In this paper, we extend the maximum likelihood approach proposed in our previous work [1] to the multi-objective learning for DNN-based speech enhancement (ML-MOL-DNN) to achieve the automatic adjustment of the dynamic range of prediction error values on different targets. The conditional likelihood function to be maximized is derived under the generalized Gaussian distribution (GGD) error model. Moreover, the control of the dynamic range of the prediction error values on different targets is achieved by the scale factors in GGD. Furthermore, we propose a method to update the shape factors automatically utilizing the one-to-one mapping between the kurtosis and shape factor in GGD instead of manual adjustment. The experimental results show that our ML-MOL-DNN can achieve better performance than MMSE-MOL-DNN in terms of different objective measures.

In this paper, we design a novel front-end processing system for speaker diarization under realistic conditions with challenging background noises. To cope with diversified environments, we first extend our perviously proposed progressive learning based speech enhancement model by adding multi-task learning in each intermediate layer. The corresponding progressive multi-target (PMT) in various layers includes both progressive ratio mask (PRM) and progressively enhanced log-power spectra (PELPS) with specified signal-to-noise ratios (SNRs). Speech distortions are commonly introduced during the front-end processing, which often deteriorate the back-end performance. However, the proposed speech enhancement model can be regarded as a bagging of models with multiple learning objectives, which provides flexibility for selecting the most appropriate output for robust speaker diarzation. In addition, a global SNR estimation is performed using the results of deep neural network (DNN) based speech activity detection (SAD) to decide whether the audio should be enhanced. We evaluate the speaker diarzation performance on the second DIHARD dataset which includes several different realistic conditions. Compared with the original data, experiments demonstrate that the enhanced data processed by our proposed method can effectively avoid the performance loss of every single domain, and achieve consistent improvements in most domains.

In our previous work, a progressive learning framework for long short-term memory (LSTM)-based speech enhancement was proposed to improve the performance in low SNR environment, where each LSTM layer is guided to learn an intermediate target with a specific SNR gain via the MMSE criterion. However, the constraint relationship among these targets is not considered in the objective function. In this paper, we incorporate two kinds of geometric constraints among these targets into the objective function to help LSTM achieve better training. One constraint is edge constraint and the other is the centroid constraint. In addition, we propose a method for constructing the intermediate targets online. It saves device storage space and alleviates the trouble of manually constructing intermediate targets. Experiment results demonstrate these geometric constraints can bring remarkable improvements in low SNR environments.

In recent years, the deep learning-based approaches are popular in the field of singe-channel speech enhancement. Convolutional neural networks (CNNs) are a standard component of many current speech enhancement system. In this study, we design a new Fully CNN (FCNN)-based regression model, which can directly achieve the 2-dimensional (2D) noisy lpg-power spectra (LPS) input to 2dimensional (2D) time-frequency mask output mapping, denoted as 2D-RFCNN. First, the whole 2D noisy LPS of one utterance is directly used as network input to make sure each convolutional filter can see more contextual information. Second, we only use the pooling operation on the frequency bin to ensure that the final dimension of frequency bin has a value of 1 and make the number of feature mapping same to frequency dimension, simultaneously. Finally, we also use the deep convolutional layers with a small size of filter, which is popularly used in speech recognition, for speech enhancement. Experiments of the CHiME-4 challenge task shows that our proposed 2D-RFCNN model not only improves the speech quality (PESQ) and intelligibility (STOI), but also reduces the recognition error rate on real test set.

In this paper, we design a novel joint framework of speech enhancement and speech separation for child speech extraction in realistic conditions, targeting the problem of extracting child speech from daily conversations in BabyTrain mega corpus. To the best of our knowledge, it is the first discussion of a feasible method for child speech extraction in realistic conditions. First, we make detailed analysis of the BabyTrain mega corpus, which is recorded in adverse environments. We observe problems of background noises, reverberations and child speech that is partially obscured by adult speech (for instance due to speaker overlap but also imitation by the adult). Motivated by this, we conduct a joint framework of speech enhancement and speech separation for child speech extraction. To measure the extraction results in realistic conditions, we propose several objective measurements to evaluate the performance of the our system, which is different from those commonly used for simulation data. Compared with the unprocessed approach and classification approach, our proposed approach can yield the best performance among all subsets of BabyTrain.

Recent encoder-decoder approaches typically employ string decoders to convert images into serialized strings for image-to-markup. However, for tree-structured representational markup, string representations can hardly cope with the structural complexity. In this work, we first show via a set of toy problems that string decoders struggle to decode tree structures, especially as structural complexity increases, we then propose a tree-structured decoder that specifically aims at generating a tree-structured markup. Our decoders works sequentially, where at each step a child node and its parent node are simultaneously generated to form a sub-tree. This sub-tree is consequently used to construct the final tree structure in a recurrent manner. Key to the success of our tree decoder is twofold, (i) it strictly respects the parent-child relationship of trees, and (ii) it explicitly outputs trees as oppose to a linear string. Evaluated on both math formula recognition and chemical formula recognition, the proposed tree decoder is shown to greatly outperform strong string decoder baselines.

Chinese character recognition has attracted much interest due to its high challenge and various applications. The whole-character modeling method can recognize common characters well but unable to handle unseen situation. Some radical-based modeling methods have successfully achieved great performance in unseen condition but need RNN-based decoder for sequence decoding. Therefore, a compact model which can recognize unseen characters needs to be proposed. First, this paper introduces a novel radical counter network (RCN) to recognize Chinese characters by identifying radicals and spatial structures. The proposed RCN first extracts visual features from input by employing DenseNet as encoder. Then a decoder based on fully connected layer is employed, aiming at synchronously estimating the number of each caption in character. Additionally, we design a multi-task learning to combine global feature extraction capability of whole-character modeling and local feature extraction capability of radical-based modeling, which further improves the model generalization. Experiments on natural scene character dataset demonstrate that the proposed model significantly outperforms WCN by 5.48% and achieve comparable performance with RAN in lower model complexity. That shows great robustness and simplicity of our model.

Recently, many researches propose to employ attention based encoder-decoder models to convert a sequence of trajectory points into a LaTeX string for online handwritten mathematical expression recognition (OHMER), and the recognition performance of these models critically relies on the accuracy of the attention. In this paper, unlike previous methods which basically employ a soft attention model, we propose to employ a posterior attention model, which modifies the attention probabilities after observing the output probabilities generated by the soft attention model. In order to further improve the posterior attention mechanism, we propose a stroke average pooling layer to aggregate point-level features obtained from the encoder into stroke-level features. We argue that posterior attention is better to be implemented on stroke-level features than point-level features as the output probabilities generated by stroke is more convincing than generated by point, and we prove that through experimental analysis. Validated on the CROHME competition task, we demonstrate that stroke based posterior attention achieves expression recognition rates of 54.26% on CROHME 2014 and 51.75% on CROHME 2016. According to attention visualization analysis, we empirically demonstrate that the posterior attention mechanism can achieve better alignment accuracy than the soft attention mechanism.

Recently, a novel radical analysis network (RAN) has the capability of effectively recognizing unseen Chinese character classes and largely reducing the requirement of training data by treating a Chinese character as a hierarchical composition of radicals rather than a single character class. However, when dealing with more challenging issues, such as the recognition of complicated characters, low-frequency character categories, and characters in natural scenes, RAN still has a lot of room for improvement. In this paper, we explore options to further improve the structure generalization and robustness capability of RAN with the Transformer architecture, which has achieved start-of-the-art results for many sequence-to-sequence tasks. More specifically, we propose to replace the original attention module in RAN with the transformer decoder, which is named as a transformer-based radical analysis network (RTN). The experimental results show that the proposed approach can significantly outperform the RAN on both printed Chinese character database and natural scene Chinese character database. Meanwhile, further analysis proves that RTN can be better generalized to complex samples and low-frequency characters, and has better robustness in recognizing Chinese characters with different attributes.

In this paper, adaptive mechanisms are applied in deep neural network (DNN) training for x-vector-based text-independent speaker verification.First, adaptive convolutional neural networks (ACNNs) are employed in frame-level embedding layers, where the parameters of the convolution filters are adjusted based on the input features.Compared with conventional CNNs, ACNNs have more flexibility in capturing speaker information.Moreover, we replace conventional batch normalization (BN) with adaptive batch normalization (ABN).By dynamically generating the scaling and shifting parameters in BN, ABN adapts models to the acoustic variability arising from various factors such as channel and environmental noises.Finally, we incorporate these two methods to further improve performance.Experiments are carried out on the speaker in the wild (SITW) and VOiCES databases.The results demonstrate that the proposed methods significantly outperform the original xvector approach.

In this paper, we propose a sub-utterance unit selection framework to remove acoustic segments in audio recordings that carry little information for acoustic scene classification (ASC). Our approach is built upon a universal set of acoustic segment units covering the overall acoustic scene space. First, those units are modeled with acoustic segment models (ASMs) used to tokenize acoustic scene utterances into sequences of acoustic segment units. Next, paralleling the idea of stop words in information retrieval, stop ASMs are automatically detected. Finally, acoustic segments associated with the stop ASMs are blocked, because of their low indexing power in retrieval of most acoustic scenes. In contrast to building scene models with whole utterances, the ASM-removed sub-utterances, i.e., acoustic utterances without stop acoustic segments, are then used as inputs to the AlexNet-L back-end for final classification. On the DCASE 2018 dataset, scene classification accuracy increases from 68%, with whole utterances, to 72.1%, with segment selection. This represents a competitive accuracy without any data augmentation, and/or ensemble strategy. Moreover, our approach compares favourably to AlexNet-L with attention.

In this paper, we exploit the properties of mean absolute error (MAE) as a loss function for the deep neural network (DNN) based vector-to-vector regression. The goal of this work is two-fold: (i) presenting performance bounds of MAE, and (ii) demonstrating new properties of MAE that make it more appropriate than mean squared error (MSE) as a loss function for DNN based vector-to-vector regression. First, we show that a generalized upper-bound for DNN-based vector-to-vector regression can be ensured by leveraging the known Lipschitz continuity property of MAE. Next, we derive a new generalized upper bound in the presence of additive noise. Finally, in contrast to conventional MSE commonly adopted to approximate Gaussian errors for regression, we show that MAE can be interpreted as an error modeled by Laplacian distribution. Speech enhancement experiments are conducted to corroborate our proposed theorems and validate the performance advantages of MAE over MSE for DNN based regression.

In our previous work, we introduced our attention-based speaker adaptation method, which has been proved to be an efficient online speaker adaptation method for real-time speech recognition. In this paper, we present a more complete framework of this method named memory-aware networks, which consists of the main network, the memory module, the attention module and the connection module. A gate mechanism and a multiple-connections strategy are presented to connect the memory with the main network in order to take full advantage of the memory. An auxiliary speaker classification task is provided to improve the accuracy of the attention module. The fixed-size ordinally forgetting encoding method is used together with average pooling to gather both short-term and long-term information. Furthermore, instead of only using traditional speaker embeddings such as i-vectors or d-vectors as the memory, we design a new form of memory called residual vectors, which can represent different pronunciation habits. Experiments on both the Switchboard and AISHELL-2 tasks show that our method can perform online speaker adaptation very well with no additional adaptation data and with only a relative 3% increase in decoding computation complexity. Under the cross-entropy criterion, our method achieves a relative word error rate reduction of 9.4% and 8.3% compared to that of the speaker-independent model on the Switchboard task and the AISHELL-2 task, respectively, and approximately 7.0% compared to that of the traditional d-vector-based speaker adaptation method.

We propose a multi-target, signal-to-noise-ratio (SNR)-progressive learning (SNR-PL) framework for regression based speech enhancement (SE). At low SNR levels, it is often not easy to directly learn the complicated regression required in SE. We therefore decompose the original SE problem of mapping noisy to clean speech features, with a large SNR gap, into a series of sub-problems, each with a small SNR increment and presumably easier to learn. In our configurations, each hidden layer of the proposed regression neural network is guided to explicitly learn an intermediate target with a specified but small SNR gain. Tested on both deep neural network (DNN) and long short-term memory (LSTM) architectures, SNR-PL consistently outperforms the conventional “black box” DNN framework in terms of both objective measure superiority and network model compactness. Furthermore, with the best configured LSTM-based SNR-PL model, we often observe that the performance is easily saturated or even degraded when increasing the number of intermediate targets, due to the fact that useful information is lost in dimension reduction when involving more target layers. Accordingly, to address this information loss issue, we explore densely connected networks on top of the LSTM structure where the input and the preceding intermediate targets are concatenated together to learn the next target. Finally, to fully utilize the rich and complementary information of intermediate targets, a simple post-processing strategy is adopted to further improve the performance. Evaluated on the simulation speech data, experimental results in unseen noises cases demonstrate that the proposed approach consistently performs better than the conventional LSTM approach in terms of objective speech enhancement measures for speech intelligibility and quality. Furthermore, when evaluated on real data provided by the CHiME-4 Challenge for automatic speech recognition (ASR) of noisy microphone array speech, we show that the proposed approach with intermediate outputs can directly improve the ASR performance, while the conventional LSTM approach increases the word error rate.

We propose a novel method for representing oriented objects in aerial images named adaptive period embedding (APE). Although traditional object detection methods represent objects using horizontal bounding boxes, the objects in aerial images are oriented. Calculating the angle of the object is a yet challenging task. Almost all previous object detectors for aerial images directly regress the angle of objects, they use complex rules to calculate the angle, and their performance is limited by the rule design. In contrast, our method is based on the angular periodicity of oriented objects. The angle is represented by two 2-D periodic vectors the periods of which are different, so the vector is continuous as the shape changes. The label generation rule is simpler and more reasonable compared with previous methods. The proposed method is general and can be applied to other oriented detector. Besides, we propose a novel intersection over union (IoU) calculation method for long objects named length-independent IoU (LIIoU). We intercept part of the long side of the target box to get the maximum IoU between the proposed box and intercepted target box. Thereby, some long boxes will have corresponding positive samples. Our method reaches the first place of DOAI2019 competition task1 (oriented object) held in a workshop on detecting objects in aerial images in conjunction with IEEE CVPR 2019.

In this paper, we design a two-stage single-channel speaker-dependent speech separation approach for the CHiME-5 Challenge, targeting the problem of far-field and multi-talker conversational speech recognition in dinner party scenarios involving background noises, reverberations and overlapping speech. First, we make detailed analysis of the CHiME-5 data and observe problems of inaccurate human annotations and low-resource useable data for target speakers. Motivated by this, we conduct a first-stage speaker-dependent speech separation with a learning target for aggressive segregation to generate more and purer target speech data. Then a second-stage speaker-dependent speech separation with a new learning target is performed to obtain the final speech masks, which can be directly fed to back-end acoustic model. Compared with the official baseline, our proposed approach can yield an absolute word error rate reduction of 5.3%, namely from 81.3% to 76.0% in development test set. To the best of our knowledge, it is the first time to discuss a feasible method of single-channel speaker-dependent speech separation for such a challenging task although we make an assumption of oracle speaker diarization following the challenge rules. By integrating this crucial technique, our submitted systems achieved the first place of all four tasks in the CHiME-5 challenge.

For conventional single-channel speech enhancement based on noise power spectrum, the speech gain function, which suppresses background noise at each time-frequency bin, is calculated by prior signal-to-noise-ratio (SNR). Hence, accurate prior SNR estimation is paramount for successful noise suppression. Accordingly, we have proposed a single-channel approach to combine conventional and deep learning techniques for speech enhancement and automatic speech recognition (ASR) recently. However, the combination process is at the testing stage, which is time-consuming with a complicated procedure. In this study, the gain function of classic speech enhancement will be utilized to optimize the ideal ratio mask based deep neural network (DNN-IRM) at the training stage, denoted as GF-DNN-IRM. And at the testing stage, the estimated IRM by GF-DNN-IRM model is directly used to generate enhanced speech without involving the conventional speech enhancement process. In addition, DNNs with less parameters in the causal processing mode are also discussed. Experiments of the CHiME-4 challenge task show that our proposed algorithm can achieve a relative word error rate reduction of 6.57% on RealData test set comparing to unprocessed speech without acoustic model retraining in causal mode, while the traditional DNN-IRM method fails to improve ASR performance in this case.

In this paper, we propose a novel multi-modal attention network (MAN), which is based on encoder-decoder framework, for handwritten mathematical expression recognition (HMER). Here, multi-modal means two specific modalities: online and offline, where online modality employs dynamic trajectories as input and offline modality employs static images as input. More specifically, the proposed method first feeds dynamic trajectories and static images into online and offline channels of the multi-modal encoder respectively. The output of the encoder is then transferred to the multi-modal decoder to generate a LaTeX sequence as the mathematical expression recognition result. To make full use of the complementary information that comes from the two modalities, we propose a re-attention mechanism as an enhanced version of the multi-modal attention mechanism which can further improve the recognition performance. Evaluated on a benchmark published by CROHME competition, the proposed approach achieves an expression recognition accuracy of 54.05% on CROHME 2014 and 50.56% on CROHME 2016 which substantially outperforms the state-of-the-arts using the single online or offline modality.

The audio-video based emotion recognition aims to classify a given video into basic emotions. In this paper, we describe our approaches in EmotiW 2019, which mainly explores emotion features and feature fusion strategies for audio and visual modality. For emotion features, we explore audio feature with both speech-spectrogram and Log Mel-spectrogram and evaluate several facial features with different CNN models and different emotion pretrained strategies. For fusion strategies, we explore intra-modal and cross-modal fusion methods, such as designing attention mechanisms to highlights important emotion feature, exploring feature concatenation and factorized bilinear pooling (FBP) for cross-modal feature fusion. With careful evaluation, we obtain 65.5% on the AFEW validation set and 62.48% on the test set and rank third in the challenge.

Automatic emotion recognition (AER) is a challenging task due to the abstract concept and multiple expressions of emotion. Although there is no consensus on a definition, human emotional states usually can be apperceived by auditory and visual systems. Inspired by this cognitive process in human beings, it's natural to simultaneously utilize audio and visual information in AER. However, most traditional fusion approaches only build a linear paradigm, such as feature concatenation and multi-system fusion, which hardly captures complex association between audio and video. In this paper, we introduce factorized bilinear pooling (FBP) to deeply integrate the features of audio and video. Specifically, the features are selected through the embedded attention mechanism from respective modalities to obtain the emotion-related regions. The whole pipeline can be completed in a neural network. Validated on the AFEW database of the audio-video sub-challenge in EmotiW2018, the proposed approach achieves an accuracy of 62.48%, outperforming the state-of-the-art result.

This paper introduces the second DIHARD challenge, the second in a series of speaker diarization challenges intended to improve the robustness of diarization systems to variation in recording equipment, noise conditions, and conversational domain. The challenge comprises four tracks evaluating diarization performance under two input conditions (single channel vs. multi-channel) and two segmentation conditions (diarization from a reference speech segmentation vs. diarization from scratch). In order to prevent participants from overtuning to a particular combination of recording conditions and conversational domain, recordings are drawn from a variety of sources ranging from read audiobooks to meeting speech, to child language acquisition recordings, to dinner parties, to web video. We describe the task and metrics, challenge design, datasets, and baseline systems for speech enhancement, speech activity detection, and diarization.

The x-vector based deep neural network (DNN) embedding systems have demonstrated effectiveness for text-independent speaker verification.This paper presents a multi-task learning architecture for training the speaker embedding DNN with the primary task of classifying the target speakers, and the auxiliary task of reconstructing the first-and higher-order statistics of the original input utterance.The proposed training strategy aggregates both the supervised and unsupervised learning into one framework to make the speaker embeddings more discriminative and robust.Experiments are carried out using the NIST SRE16 evaluation dataset and the VOiCES dataset.The results demonstrate that our proposed method outperforms the original x-vector approach with very low additional complexity added.

In this paper, gating mechanisms are applied in deep neural network (DNN) training for x-vector-based text-independent speaker verification. First, a gated convolution neural network (GCNN) is employed for modeling the frame-level embedding layers. Compared with the time-delay DNN (TDNN), the GCNN can obtain more expressive frame-level representations through carefully designed memory cell and gating mechanisms. Moreover, we propose a novel gated-attention statistics pooling strategy in which the attention scores are shared with the output gate. The gated-attention statistics pooling combines both gating and attention mechanisms into one framework; therefore, we can capture more useful information in the temporal pooling layer. Experiments are carried out using the NIST SRE16 and SRE18 evaluation datasets. The results demonstrate the effectiveness of the GCNN and show that the proposed gated-attention statistics pooling can further improve the performance.

From a statistical perspective, the conventional minimum mean squared error (MMSE) criterion can be considered as the maximum likelihood (ML) solution under an assumed homoscedastic Gaussian error model. However, in this paper, a statistical analysis reveals the super-Gaussian and heteroscedastic properties of the prediction errors in nonlinear regression deep neural network (DNN)-based speech enhancement when estimating clean log-power spectral (LPS) components at DNN outputs with noisy LPS features in DNN input vectors. Accordingly, we propose treating all dimensions of the prediction error vector as statistically independent random variables and model them with generalized Gaussian distributions (GGDs). Then, the objective function with the GGD error model is derived according to the ML criterion. Experiments on the TIMIT corpus corrupted by simulated additive noises show consistent improvements of our proposed DNN framework over the conventional DNN framework in terms of various objective quality measures under 14 unseen noise types evaluated and at various signal-to-noise ratio levels. Furthermore, the ML optimization objective with GGD outperforms the conventional MMSE criterion, achieving improved generalization and robustness.

Automatic speech recognition (ASR) systems are often built using scene related speech data due to large variations of transmission channels and sampling rates in different scenarios. In this study, we propose a general framework that establishes a unified model for diversified speech data with different sampling rates and channels. The framework is a joint optimization of deep neural network (DNN)-based bandwidth expansion and acoustic modeling to exploit a large amount of diversified training data. First, we design two novel DNN architectures to map the acoustic features from narrowband to wideband speech through direct mapping and progressive mapping. The learning targets of the direct mapping DNN (DNN-DM) are the acoustic features extracted from speech with the largest bandwidth, while the acoustic features from speech with all the other bandwidths are used as input. A progressive stacking network (PSN) gradually maps the features from the low sampling rates to the highest sampling rate through the design of intermediate target layers via multitask training. Then, in addition to these bandwidth expansion networks, we investigate several joint training strategies for DNN-based acoustic models. Our experiments conducted on three diversified large-scale Mandarin speech datasets with different recording channels and sampling rates (6, 8, and 16 kHz) show that the proposed unified model using PSN for bandwidth expansion not only is a more flexible and compact design than conventional multiple acoustic models with each bandwidth for a specific sampling rate, but also yields consistent and significant improvements over bandwidth-dependent models with an average relative word error rate reduction of 6.2%, indicating that the proposed model can fully utilize the diversified cross-channel speech data with multiple bandwidths. Moreover, the proposed methods are verified to be robust on different realistic scenes and can be effectively extended to a long short-term memory framework.

This paper focuses on a theoretical analysis of deep neural network (DNN) based functional approximation. Leveraging upon two classical theorems on universal approximation, an artificial neural network (ANN) with a single hidden layer of neurons is used. With modified ReLU and Sigmoid activation functions, we first generalize the related concepts to vector-to-vector regression. Then, we show that the width of the hidden layer of ANN is numerically related to the approximation of the regression function. Furthermore, we increase the number of hidden layers and show that the depth of the ANN-based regression function can enhance its expressive power. We illustrate this representation with recently-emerged DNN based speech enhancement. We first compare the expressive power by varying ANN structures and then test its related regression performance under different noisy conditions in various noise types and signal-to-noise-ratio levels. Experimental results verify our theoretical prediction that an ANN of a broader hidden layer and a deeper architecture can jointly ensure a closer approximation of the vector-to-vector regression functions in terms of the Euclidean distance between the log power spectra of noisy and expected clean speech. Moreover, a DNN with a broader width at the top hidden layer can improve the regression performance relative to those with a narrower width at the top hidden layers.

In this paper, we propose a novel teacher-student learning framework for the preprocessing of a speech recognizer, leveraging the online noise tracking capabilities of improved minima controlled recursive averaging (IMCRA) and deep learning of nonlinear interactions between speech and noise. First, a teacher model with deep architectures is built to learn the target of ideal ratio masks (IRMs) using simulated training pairs of clean and noisy speech data. Next, a student model is trained to learn an improved speech presence probability by incorporating the estimated IRMs from the teacher model into the IMCRA approach. The student model can be compactly designed in a causal processing mode having no latency with the guidance of a complex and noncausal teacher model. Moreover, the clean speech requirement, which is difficult to meet in real-world adverse environments, can be relaxed for training the student model, implying that noisy speech data can be directly used to adapt the regression-based enhancement model to further improve speech recognition accuracies for noisy speech collected in such conditions. Experiments on the CHiME-4 challenge task show that our best student model with bidirectional gated recurrent units (BGRUs) can achieve a relative word error rate (WER) reduction of 18.85% for the real test set when compared to unprocessed system without acoustic model retraining. However, the traditional teacher model degrades the performance of the unprocessed system in this case. In addition, the student model with a deep neural network (DNN) in causal mode having no latency yields a relative WER reduction of 7.94% over the unprocessed system with 670 times less computing cycles when compared to the BGRU-equipped student model. Finally, the conventional speech enhancement and IRM-based deep learning method destroyed the ASR performance when the recognition system became more powerful. While our proposed approach could still improve the ASR performance even in the more powerful recognition system.

In this paper, we introduce Track, Attend, and Parse (TAP), an end-to-end approach based on neural networks for online handwritten mathematical expression recognition (OHMER). The architecture of TAP consists of a tracker and a parser. The tracker employs a stack of bidirectional recurrent neural networks with gated recurrent units (GRU) to model the input handwritten traces, which can fully utilize the dynamic trajectory information in OHMER. Followed by the tracker, the parser adopts a GRU equipped with guided hybrid attention (GHA) to generate notations. The proposed GHA is composed of a coverage-based spatial attention, a temporal attention, and an attention guider. Moreover, we demonstrate the strong complementarity between offline information with static-image input and online information with ink-trajectory input by blending a fully convolutional networks-based watcher into TAP. Inherently, unlike traditional methods, this end-to-end framework does not require the explicit symbol segmentation and a predefined expression grammar for parsing. Validated on a benchmark published by the CROHME competition, the proposed approach outperforms the state-of-the-art methods and achieves the best reported results with an expression recognition accuracy of 61.16% on CROHME 2014 and 57.02% on CROHME 2016, using only official training dataset.

Recently, we proposed a novel progressive learning (PL) framework for deep neural network (DNN) based speech enhancement to improve the performance in low signal-to-noise ratio (SNR) environments. In this study, several new contributions are made to this framework. First, the advanced long short-term memory (LSTM) architecture is adopted to achieve better results, namely LSTM-PL, where each LSTM layer is guided to explicitly learn an intermediate target with a specific SNR gain. However, we observe that the performance of LSTM-PL architecture is easily degraded by increasing the number of intermediate targets due to the possible information loss when involving more target layers. Accordingly, we propose densely connected progressive learning in which the input and the estimations of intermediate targets are spliced together to learn the next target. This new structure can fully utilize the rich set of information from the multiple learning targets and alleviate the information loss problem. Experimental results demonstrate that the dense structure with deeper LSTM layers can yield significant gains of speech intelligibility measure for all noise types and levels. Moreover, the post-processing with more targets tends to achieve better performance.

Automatic speech recognition is more and more widely and effectively used. Nevertheless, in some automatic speech analysis tasks the state of the art is surprisingly poor. One of these is “diarization”, the task of determining who spoke when. Diarization is key to processing meeting audio and clinical interviews, extended recordings such as police body cam or child language acquisition data, and any other speech data involving multiple speakers whose voices are not cleanly separated into individual channels. Overlapping speech, environmental noise and suboptimal recording techniques make the problem harder. During the JSALT Summer Workshop at CMU in 2017, an international team of researchers worked on several aspects of this problem, including calibration of the state of the art, detection of overlaps, enhancement of noisy recordings, and classification of shorter speech segments. This paper sketches the workshop's results, and announces plans for a “Diarization Challenge” to encourage further progress.

In this study, we investigate on the effects of deep learning based speech enhancement as a preprocessor to speaker diarization in quite challenging realistic environments involving the background noises, reverberations and overlapping speech. To improve the generalization capability, the advanced long short-term memory (LSTM) architecture with the novel design of hidden layers via densely connected progressive learning and output layer via multiple-target learning is proposed for preprocessing. We build the deep model using synthesized training data pairs generated from WSJO reading-style speech and more than 100 noise types. Surprisingly, this proposed preprocessor demonstrates a strong generalization capability to speaker di-arization with the realistic noisy speech in highly mismatched conditions, in terms of the speaking style, interferences, and the interaction between them. Tested on three challenging tasks, namely AMI, ADOS, and SeedLings, the state-of-the-art diarization system with the novel LSTM-based speech preprocessor can yield consistent and significant reductions of diarization error rate (DER) over the systems using unprocessed noisy speech and traditional enhancement methods.

Recently, hidden Markov models (HMMs) have achieved promising results for offline handwritten Chinese text recognition. However, due to the large vocabulary of Chinese characters with each modeled by a uniform and fixed number of hidden states, a high demand of memory and computation is required. In this study, to address this issue, we present parsimonious HMMs via the state tying which can fully utilize the similarities among different Chinese characters. Two-step algorithm with the data-driven question-set is adopted to generate the tied-state pool using the likelihood measure. The proposed parsimonious HMMs with both Gaussian mixture models (GMMs) and deep neural networks (DNNs) as the emission distributions not only lead to a compact model but also improve the recognition accuracy via the data sharing for the tied states and the confusion decreasing among state classes. Tested on ICDAR-2013 competition database, in the best configured case, the new parsimonious DNN-HMM can yield a relative character error rate (CER) reduction of 6.2%, 25% reduction of model size and 60% reduction of decoding time over the conventional DNN-HMM. In the compact setting case of average 1-state HMM, our parsimonious DNN-HMM significantly outperforms the conventional DNN-HMM with a relative CER reduction of 35.5%.

Recently, great success has been achieved in offline handwritten Chinese character recognition by using deep learning methods. Chinese characters are mainly logographic and consist of basic radicals, however, previous research mostly treated each Chinese character as a whole without explicitly considering its internal two-dimensional structure and radicals. In this study, we propose a novel radical analysis network with densely connected architecture (DenseRAN) to analyze Chinese character radicals and its two-dimensional structures simultaneously. DenseRAN first encodes input image to high-level visual features by employing DenseNet as an encoder. Then a decoder based on recurrent neural networks is employed, aiming at generating captions of Chinese characters by detecting radicals and two-dimensional structures through attention mechanism. The manner of treating a Chinese character as a composition of two-dimensional structures and radicals can reduce the size of vocabulary and enable DenseRAN to possess the capability of recognizing unseen Chinese character classes, only if the corresponding radicals have been seen in training set. Evaluated on ICDAR-2013 competition database, the proposed approach significantly outperforms whole-character modeling approach with a relative character error rate (CER) reduction of 18.54%. Meanwhile, for the case of recognizing 3277 unseen Chinese characters in CASIA-HWDB1.2 database, DenseRAN can achieve a character accuracy of about 41% while the traditional whole-character method has no capability to handle them.

Chinese characters have a huge set of character categories, more than 20, 000 and the number is still increasing as more and more novel characters continue being created. However, the enormous characters can be decomposed into a compact set of about 500 fundamental and structural radicals. This paper introduces a novel radical analysis network (RAN) to recognize printed Chinese characters by identifying radicals and analyzing two-dimensional spatial structures among them. The proposed RAN first extracts visual features from input by employing convolutional neural networks as an encoder. Then a decoder based on recurrent neural networks is employed, aiming at generating captions of Chinese characters by detecting radicals and two-dimensional structures through a spatial attention mechanism. The manner of treating a Chinese character as a composition of radicals rather than a single character class largely reduces the size of vocabulary and enables RAN to possess the ability of recognizing unseen Chinese character classes, namely zero-shot learning.

Handwritten mathematical expression recognition is a challenging problem due to the complicated two-dimensional structures, ambiguous handwriting input and variant scales of handwritten math symbols. To settle this problem, recently we propose the attention based encoder-decoder model that recognizes mathematical expression images from two-dimensional layouts to one-dimensional LaTeX strings. In this study, we improve the encoder by employing densely connected convolutional networks as they can strengthen feature extraction and facilitate gradient propagation especially on a small training set. We also present a novel multi-scale attention model which is employed to deal with the recognition of math symbols in different scales and restore the fine-grained details dropped by pooling operations. Validated on the CROHME competition task, the proposed method significantly outperforms the state-of-the-art methods with an expression recognition accuracy of 52.8% on CROHME 2014 and 50.1% on CROHME 2016, by only using the official training dataset.

Recently, great progress has been made for online handwritten Chinese character recognition due to the emergence of deep learning techniques. However, previous research mostly treated each Chinese character as one class without explicitly considering its inherent structure, namely the radical components with complicated geometry. In this study, we propose a novel trajectory-based radical analysis network (TRAN) to firstly identify radicals and analyze two-dimensional structures among radicals simultaneously, then recognize Chinese characters by generating captions of them based on the analysis of their internal radicals. The proposed TRAN employs recurrent neural networks (RNNs) as both an encoder and a decoder. The RNN encoder makes full use of online information by directly transforming handwriting trajectory into high-level features. The RNN decoder aims at generating the caption by detecting radicals and spatial structures through an attention model. The manner of treating a Chinese character as a two-dimensional composition of radicals can reduce the size of vocabulary and enable TRAN to possess the capability of recognizing unseen Chinese character classes, only if the corresponding radicals have been seen. Evaluated on CASIA-OLHWDB database, the proposed approach significantly outperforms the state-of-the-art whole-character modeling approach with a relative character error rate (CER) reduction of 10%. Meanwhile, for the case of recognition of 500 unseen Chinese characters, TRAN can achieve a character accuracy of about 60 % while the traditional whole-character method has no capability to handle them.

Traditional text detection methods mostly focus on quadrangle text. In this study we propose a novel method named sliding line point regression (SLPR) in order to detect arbitrary-shape text in natural scene. SLPR regresses multiple points on the edge of text line and then utilizes these points to sketch the outlines of the text. The proposed SLPR can be adapted to many object detection architectures such as Faster R-CNN and R-FCN. Specifically, we first generate the smallest rectangular box including the text with region proposal network (RPN), then isometrically regress the points on the edge of text by using the vertically and horizontally sliding lines. To make full use of information and reduce redundancy, we calculate x-coordinate or y-coordinate of target point by the rectangular box position, and just regress the remaining y-coordinate or x-coordinate. Accordingly we can not only reduce the parameters of system, but also restrain the points which will generate more regular polygon. Our approach achieved competitive results on traditional ICDAR2015 Incidental Scene Text benchmark and curve text detection dataset CTW1500.

In this study, we propose a novel deep neural network (DNN) architecture for speech enhancement (SE) via a multiobjective learning and ensembling (MOLE) framework to achieve a compact and lowlatency design, while maintaining good performance in quality evaluations. MOLE follows the boosting concept when combining weak models into a strong classifier and consists of two compact DNNs. The first, called the multiobjective learning DNN (MOL-DNN), takes multiple features, such as log-power spectra (LPS), mel-frequency cepstral coefficients (MFCCs) and Gammatone frequency cepstral coefficients (GFCCs) to predict a multiobjective set that includes clean speech feature, dynamic noise feature, and ideal ratio mask (IRM). The second, called the multiobjective ensembling DNN (MOE-DNN), takes the learned features from MOL-DNN as inputs and separately predicts clean LPS and IRM, clean MFCC and IRM, and clean GFCC and IRM using three sets of weak regression functions. Finally, a postprocessing operation can be applied to the estimated clean features by leveraging the multiple targets learned from both the MOL-DNN and the MOE-DNN. On speech corrupted by 15 noise types not seen in model training the SE results show that the MOLE approach, which features a small model size and low run-time latency, can achieve consistent improvements over both DNN- and long short-term memory (LSTM)-based techniques in terms of all the objective metrics evaluated in this study for all three cases (the input contexts contain 1-frame, 4-frame and 7-frame instances). The 1-frame MOLE-based SE system outperforms the DNN-based SE system with a 7-frame input expansion at a 3-frame delay and also achieves better performance than the LSTM-based SE system with 4-frame, no delay expansion by including only 3 previous frames, and with 170 times less processing latency.

In this study, we present a novel end-to-end approach based on the encoder-decoder framework with the attention mechanism for online handwritten mathematical expression recognition (OHMER). First, the input two-dimensional ink trajectory information of handwritten expression is encoded via the gated recurrent unit based recurrent neural network (GRU-RNN). Then the decoder is also implemented by the GRU-RNN with a coverage-based attention model. The proposed approach can simultaneously accomplish the symbol recognition and structural analysis to output a character sequence in LaTeX format. Validated on the CROHME 2014 competition task, our approach significantly outperforms the state-of-the-art with an expression recognition accuracy of 52.43% by only using the official training dataset. Furthermore, the alignments between the input trajectories of handwritten expressions and the output LaTeX sequences are visualized by the attention mechanism to show the effectiveness of the proposed method.

In this paper, we investigate high-resolution modeling units of deep neural networks (DNNs) from concrete to abstract for acoustic scene classification based on Gaussian mixture model (GMM) and ergodic hidden Markov model (HMM). A direct modeling strategy for DNN to classify acoustic scenes is to map each frame feature of an audio to one scene category. However, all frames tagged with the same label may not be the best choice because the representative pattern of an audio is sparse. GMM is also often employed to model each acoustic scene directly as a generative model. Because the multiple Gaussians in a GMM model have different levels of contribution, and each Gaussian can be seen as a subclass of the scene category, so we can utilize the subclass of GMM as a bit abstract modeling unit to adopt DNN-GMM system. When single scene category is subdivided into various subclasses, prior scores for each subclass calculated from training set are stored as one part of model to response the sparseness of representative pattern. Ergodic HMM should be more appropriate to model the acoustic scenes than GMM due to the uncertain structure of scene audio. Using HMM states as modeling units, we build DNN-HMM hybrid system. By comparison, we find high-resolution modeling units are more effective than direct modeling. The final system is obtained by performing system combination to take advantage of the complementarity of different-level modeling units. Experiments on acoustic scene classification task of DCASE2016 challenge show that our final system yields 25.9% relative error rate reduction compared with a GMM baseline on evaluation set.

We propose an unsupervised speech separation framework for mixtures of two unseen speakers in a single-channel setting based on deep neural networks (DNNs). We rely on a key assumption that two speakers could be well segregated if they are not too similar to each other. A dissimilarity measure between two speakers is first proposed to characterize the separation ability between competing speakers. We then show that speakers with the same or different genders can often be separated if two speaker clusters, with large enough distances between them, for each gender group could be established, resulting in four speaker clusters. Next, a DNN-based gender mixture detection algorithm is proposed to determine whether the two speakers in the mixture are females, males, or from different genders. This detector is based on a newly proposed DNN architecture with four outputs, two of them representing the female speaker clusters and the other two characterizing the male groups. Finally, we propose to construct three independent speech separation DNN systems, one for each of the female-female, male-male, and female-male mixture situations. Each DNN gives dual outputs, one representing the target speaker group and the other characterizing the interfering speaker cluster. Trained and tested on the speech separation challenge corpus, our experimental results indicate that the proposed DNN-based approach achieves large performance gains over the state-of-the-art unsupervised techniques without using any specific knowledge about the mixed target and interfering speakers being segregated.

Recently, we propose deep neural network based hidden Markov models (DNN-HMMs) for offline handwritten Chinese text recognition. In this study, we design a novel writer code based adaptation on top of the DNN-HMM to further improve the accuracy via a customized recognizer. The writer adaptation is implemented by incorporating the new layers with the original input or hidden layers of the writer-independent DNN. These new layers are driven by the so-called writer code, which guides and adapts the DNN-based recognizer with the writer information. In the training stage, the writer-aware layers are jointly learned with the conventional DNN layers in an alternative manner. In the recognition stage, with the initial recognition results from the first-pass decoding with the writer-independent DNN, an unsupervised adaptation is performed to generate the writer code via the cross-entropy criterion for the subsequent second-pass decoding. The experiments on the most challenging task of ICDAR 2013 Chinese handwriting competition show that our proposed adaptation approach can achieve consistent and significant improvements of recognition accuracy over a highperformance writer-independent DNN-HMM based recognizer across all 60 writers, yielding a relative character error rate reduction of 23.62% in average.

This paper proposes a novel segmentation-free approach using deep neural network based hidden Markov model (DNN-HMM) for offline handwritten Chinese text recognition. In the general Bayesian framework, three key issues are comprehensively investigated, namely feature extraction, character modeling, and language modeling. First, as for the feature extraction on the basis of each frame or sliding window, the gradient-based features are extracted for the DNN-based classifier. Second, the text line is sequentially modeled by HMMs with each representing one character class. Meanwhile the DNN-based classifier is adopted to calculate the posterior probability of all HMM states. Finally, the character n-gram language model is integrated with the DNN-HMM character model for the Bayesian decision. The experiments on the ICDAR 2013 competition task of CASIA-HWDB database show that the proposed approach can achieve the best published recognition results to our knowledge, yielding a character error rate (CER) of 6.50%, which significantly outperforms the previously best reported oversegmentation approach (with a CER of 9.25%) and the segmentation-free approach using multidimensional long-short term memory recurrent neural network (MDLSTM-RNN) approach (with a CER of 10.6%).

We propose joint modeling strategies leveraging upon large-scale mixed-band training speech for recognition of both narrowband and wideband data based on deep neural networks (DNNs). We utilize conventional down-sampling and up-sampling schemes to go between narrowband and wideband data. We also explore DNN-based speech bandwidth expansion (BWE) to map some acoustic features from narrowband to wideband speech. By arranging narrowband and wideband features at the input or the output level of BWE-DNN, and combining down-sampling and up-sampling data, different DNNs can be established. Our experiments on a Mandarin speech recognition task show that the hybrid DNNs for joint modeling of mixed-band speech yield significant performance gains over both the narrowband and wideband speech models, well-trained separately, with a relative character error rate reduction of 7.9% and 3.9% on narrowband and wideband data, respectively. Furthermore, the proposed strategies also consistently outperform other conventional DNN-based methods.

We propose a novel data-driven approach to single-channel speech separation based on deep neural networks (DNNs) to directly model the highly nonlinear relationship between speech features of a mixed signal containing a target speaker and other interfering speakers. We focus our discussion on a semisupervised mode to separate speech of the target speaker from an unknown interfering speaker, which is more flexible than the conventional supervised mode with known information of both the target and interfering speakers. Two key issues are investigated. First, we propose a DNN architecture with dual outputs of the features of both the target and interfering speakers, which is shown to achieve a better generalization capability than that with output features of only the target speaker. Second, we propose using a set of multiple DNNs, each intending to be signal-noise-dependent (SND), to cope with the difficulty that one single general DNN could not well accommodate all the speaker mixing variabilities at different signal-to-noise ratio (SNR) levels. Experimental results on the speech separation challenge (SSC) data demonstrate that our proposed framework achieves better separation results than other conventional approaches in a supervised or semisupervised mode. SND-DNNs could also yield significant performance improvements over a general DNN for speech separation in low SNR cases. Furthermore, for automatic speech recognition (ASR) following speech separation, this purely front-end processing with a single set of speaker-independent ASR acoustic models, achieves a relative word error rate (WER) reduction of 11.6% over a state-of-the-art separation and recognition system where a complicated joint back-end decoding framework with multiple sets of speaker-dependent ASR acoustic models needs to be implemented. When speaker-adaptive ASR acoustic models for the target speakers are adopted for the enhanced signals, another 12.1% WER reduction over our best speaker-independent ASR system is achieved.

We present an information fusion approach to robust recognition of microphone array speech for the recently launched 3rd CHiME Challenge. It is based on a deep learning framework with a large neural network consisting of subnets with different architectures. Multiple knowledge sources are integrated via an early fusion of normalized noisy features with different beamforming techniques, speech enhanced features, speaker related features, and other auxiliary features concatenated as the input to each subnet, and a late fusion by combining the outputs of all subnets to produce one single output set. Our experiments demonstrate that all information sources are complementary in our proposed framework. Our best system achieves an average word error rate reduction of 68% from the officially released baseline results on the test set of real data.

Based on the recently proposed speech pre-processing front-end with deep neural networks (DNNs), we first investigate different feature mapping directly from noisy speech via DNN for robust speech recognition. Next, we propose to jointly train a single DNN for both feature mapping and acoustic modeling. In the end, we show that the word error rate (WER) of the jointly trained system could be significantly reduced by the fusion of multiple DNN pre-processing systems which implies that features obtained from different domains of the DNN-enhanced speech signals are strongly complementary. Testing on the Aurora4 noisy speech recognition task our best system with multi-condition training can achieves an average WER of 10.3%, yielding a relative reduction of 16.3% over our previous DNN pre-processing only system with a WER of 12.3%. To the best of our knowledge, this represents the best published result on the Aurora4 task without using any adaptation techniques.

In this paper, we propose a new signal-noise-dependent (SND) deep neural network (DNN) framework to further improve the separation and recognition performance of the recently developed technique for general DNN-based speech separation. We adopt a divide and conquer strategy to design the proposed SND-DNNs with higher resolutions that a single general DNN could not well accommodate for all the speaker mixing variabilities at different levels of signal-to-noise ratios (SNRs). In this study two kinds of SNR-dependent DNNs, namely positive and negative DNNs, are trained to cover the mixed speech signals with positive and negative SNR levels, respectively. At the separation stage, a first-pass separation using a general DNN can give an accurate SNR estimation for a model selection. Experimental results on the Speech Separation Challenge (SSC) task show that SND-DNNs could yield significant performance improvements for both speech separation and recognition over a general DNN. Furthermore, this purely front-end processing method achieves a relative word error rate reduction of 11.6% over a state-of-the-art recognition system where a complicated joint decoding framework needs to be implemented in the back-end.

This paper presents a novel approach to writer adaptation based on convolutional neural network (CNN) as a feature extractor and improved discriminative linear regression for online handwritten Chinese character recognition. First, the proposed recognizer consisting of CNN-based feature extractor and prototype-based classifier can achieve comparable performance with the state-of-the-art CNN-based classifier while it could be designed more compact and efficient as a practical solution. Second, the writer adaption is performed via a linear transformation of the extracted feature from CNN. The transformation parameters are optimized with a so-called sample separation margin based minimum classification error criterion, which can be further improved by using more synthesized adaptation data and a simple regularization method. The experiments on the data collected from user inputs of Smartphones with a vocabulary of 20,936 characters demonstrate that our writer adaptation approach can yield significant improvements of recognition accuracy over a high-performance baseline system and also outperform a state-of-the-art approach based on style transfer mapping especially with increased adaptation data.

We propose a multi-objective framework to learn both secondary targets not directly related to the intended task of speech enhancement (SE) and the primary target of the clean log-power spectra (LPS) features to be used directly for constructing the enhanced speech signals.In deep neural network (DNN) based SE we introduce an auxiliary structure to learn secondary continuous features, such as mel-frequency cepstral coefficients (MFCCs), and categorical information, such as the ideal binary mask (IBM), and integrate it into the original DNN architecture for joint optimization of all the parameters.This joint estimation scheme imposes additional constraints not available in the direct prediction of LPS, and potentially improves the learning of the primary target.Furthermore, the learned secondary information as a byproduct can be used for other purposes, e.g., the IBM-based post-processing in this work.A series of experiments show that joint LPS and MFCC learning improves the SE performance, and IBM-based post-processing further enhances listening quality of the reconstructed speech.

In contrast to the conventional minimum mean square error (MMSE)-based noise reduction techniques, we propose a supervised method to enhance speech by means of finding a mapping function between noisy and clean speech signals based on deep neural networks (DNNs). In order to be able to handle a wide range of additive noises in real-world situations, a large training set that encompasses many possible combinations of speech and noise types, is first designed. A DNN architecture is then employed as a nonlinear regression function to ensure a powerful modeling capability. Several techniques have also been proposed to improve the DNN-based speech enhancement system, including global variance equalization to alleviate the over-smoothing problem of the regression model, and the dropout and noise-aware training strategies to further improve the generalization capability of DNNs to unseen noise conditions. Experimental results demonstrate that the proposed framework can achieve significant improvements in both objective and subjective measures over the conventional MMSE based technique. It is also interesting to observe that the proposed DNN approach can well suppress highly nonstationary noise, which is tough to handle in general. Furthermore, the resulting DNN model, trained with artificial synthesized data, is also effective in dealing with noisy speech data recorded in real-world scenarios without the generation of the annoying musical artifact commonly observed in conventional enhancement methods.

In our previous work, we extend the traditional stereo-based stochastic mapping by relaxing the constraint of stereo-data, which is not practical in real applications, via HMM-based speech synthesis to construct the “clean” channel data for noisy speech recognition. In this paper, we propose to use deep neural networks (DNNs) for stereo mapping compared with the joint Gaussian mixture model (GMM). The experimental results on Aurora3 databases show that our proposed DNN based synthesized stereo mapping can achieve consistently significant improvements of recognition performance over joint GMM based synthesized stereo mapping in the well-matched (WM) condition among four different European languages.

This paper presents a novel approach to writer adaptation using bottleneck features and discriminative linear regression for the recognition of online handwritten Chinese characters. First, bottleneck features extracted from a bottleneck layer of a deep neural network representing a nonlinear and discriminative transformation of the input features are verified to be much more effective in adaptation of writing styles than the conventional features after linear discriminant analysis transformation. Second, discriminative linear regression via a so-called sample separation margin based minimum classification error criterion is adopted for writer adaptation. The experiments on an in-house developed online Chinese handwriting corpus with a vocabulary of 15,167 characters and testing data collected from user inputs of Smartphones show that our proposed approach can achieve very significant improvements of recognition accuracy compared with a state-of-the-art adaptation approach for writer adaptation.

This paper presents a study of designing compact classifiers using deep neural networks for recognition of online handwritten Chinese characters. Two schemes are investigated based on practical considerations. First, deep neural networks are adopted purely as a classifier with a state-of-the-art feature extractor of online handwritten Chinese characters. Second, the so-called bottleneck features extracted from a bottleneck layer of deep neural networks are fed to the prototype-based classifier. The experiments on an in-house developed online Chinese handwriting corpus with a vocabulary of 15,167 characters show that compared with prototype-based classifier widely developed on the mobile device, deep neural network based classifier can yield significant improvements of recognition accuracy with acceptably increased footprint and latency while the bottleneck-feature approach can bring a more compact classifier with an observable performance gain.

We propose three algorithms to address the mismatch problem in deep neural network (DNN) based speech enhancement. First, we investigate noise aware training by incorporating noise informationin the testutterance with anideal binary maskbased dynamic noise estimation approach to improve DNN’s speech separation ability from the noisy signal. Next, a set of more than 100 noise types is adopted to enrich the generalization capabilities of the DNN to unseen and non-stationary noise conditions. Finally, the quality of the enhanced speech can further be improved by global variance equalization. Empirical results show that each of the three proposed techniques contributes to the performance improvement. Compared to the conventional logarithmic minimum mean squared error speech enhancement method, our DNN system achieves 0.32 PESQ (perceptual evaluation of speech quality) improvement across six signal-tonoise ratio levels ranging from -5dB to 20dB on a test set with unknown noise types. We also observe that the combined strategies can well suppress highly non-stationary noise better than all the competing state-of-the-art techniques we have evaluated. Index Terms: Speech enhancement, deep neural networks, noise aware training, ideal binary mask, non-stationary noise

This letter presents a regression-based speech enhancement framework using deep neural networks (DNNs) with a multiple-layer deep architecture. In the DNN learning process, a large training set ensures a powerful modeling capability to estimate the complicated nonlinear mapping from observed noisy speech to desired clean signals. Acoustic context was found to improve the continuity of speech to be separated from the background noises successfully without the annoying musical artifact commonly observed in conventional speech enhancement algorithms. A series of pilot experiments were conducted under multi-condition training with more than 100 hours of simulated speech data, resulting in a good generalization capability even in mismatched testing conditions. When compared with the logarithmic minimum mean square error approach, the proposed DNN-based algorithm tends to achieve significant improvements in terms of various objective quality measures. Furthermore, in a subjective preference evaluation with 10 listeners, 76.35% of the subjects were found to prefer DNN-based enhanced speech to that obtained with other conventional technique.

In our previous work, we proposed a feature compensation approach using high-order vector Taylor series (VTS) approximation for noisy speech recognition. In this paper, we report new progress on making it more powerful and practical in real applications. First, mixtures of densities are used to enhance the distortion models of both additive noise and convolutional distortion. New formulations for maximum likelihood (ML) estimation of distortion model parameters, and minimum mean squared error (MMSE) estimation of clean speech are derived and presented. Second, we improve the feature compensation in both efficiency and accuracy by applying higher order information of VTS approximation only to the noisy speech mean parameters, and a temporal smoothing operation for the posterior probability of Gaussian mixture components in clean speech estimation. Finally, we design a procedure to perform irrelevant variability normalization (IVN) based joint training of a reference Gaussian mixture model (GMM) for feature compensation and hidden Markov models (HMMs) for acoustic modeling using VTS-based feature compensation. The effectiveness of our proposed approach is confirmed by experiments on Aurora3 benchmark database for a real-world in-vehicle connected digits recognition task. Compared with ETSI advanced front-end, our approach achieves significant recognition accuracy improvement across three “training-testing” conditions for four languages.

Recently, we proposed an approach to irrelevant variability normalization (IVN) based joint training of a reference Gaussian mixture model (GMM) for feature compensation and hidden Markov models (HMMs) for acoustic modeling by using a vector Taylor series (VTS) based feature compensation technique, where single-component densities are used to model additive noise and convolutional distortion respectively. In this paper, mixtures of densities are used to enhance the distortion model. New formulations for maximum likelihood (ML) estimation of distortion model parameters, and minimum mean squared error (MMSE) estimation of clean speech are derived and presented. A comparative study is conducted under three “training-testing” conditions on Aurora3 database. Experimental results confirm that the proposed mixture models of distortion can achieve significant performance gain compared with the traditional distortion modeling.

This paper presents a discriminative training approach to irrelevant variability normalization (IVN) based joint training of feature transforms and prototype-based classifier for recognition of online handwritten Chinese characters. A sample separation margin based minimum classification error criterion is adopted in IVN-based training, while an Rprop algorithm is used for optimizing the objective function. The IVN-trained recognizer can be made both compact and efficient by using a two-level fast-match tree whose internal nodes coincide with the labels of feature transforms. The effectiveness of the proposed approach is confirmed on an online handwritten character recognition task with a vocabulary of 9,306 characters.

We present a study of designing compact multiple-prototype based classifiers for rotation-free recognition of online handwritten Chinese characters. Several versions of Rprop algorithms are adopted to optimize a sample-separation-margin based minimum classification error objective function. Split vector quantization technique is used to compress classifier parameters and a fast-match tree is used for efficient recognition. A new preprocessing technique is proposed to achieve rotation-free recognition capability. Promising benchmark results are reported on an online handwritten character recognition task with a vocabulary of 27,720 characters.

We have developed a prototype of a mobile app called "Snap and Translate" on "Windows Phone 7". A person who is reading an English menu/sign and wants a Chinese translation of an English word or phrase or paragraph can use a Windows Phone to snap an image of the text, tap the word or swipe the phrase or circle the paragraph with a finger, and get a Chinese translation displayed on the screen of the phone. This is enabled by seamless integration of three Microsoft technologies: intelligent text extraction, OCR, and machine translation based on a client-plus-cloud architecture. The current prototype also supports Chinese OCR plus Chinese-to-English translation. In this paper, we highlight the UI design of the system and the corresponding user-intention guided text extraction approach to achieving a compelling user experience.

This paper presents a new feature compensation approach to noisy speech recognition by using high-order vector Taylor series (HOVTS) approximation of an explicit model of environmental distortions. Formulations for maximum-likelihood (ML) estimation of both additive noises and convolutional distortions, and minimum mean squared error (MMSE) estimation of clean speech are derived. Experimental results on Aurora2 and Aurora4 benchmark databases, where the modeling assumption of the distortion model is more accurate, demonstrate that the standard HOVTS-based feature compensation approaches achieve consistently significant improvement in recognition accuracy compared to traditional standard first-order VTS-based approach. For a real-world in-vehicle connected digits recognition task on Aurora3 benchmark database where the modeling assumption of the distortion model is less accurate, modifications are necessary to make VTS-based feature compensation approaches work. In this case, the second-order VTS-based approach performs only slightly better than the first-order VTS-based approach.

In this paper, we apply the well-known boosted mixture learning (BML) method to learn Gaussian mixture HMMs in speech recognition. BML is an incremental method to learn mixture models for classification problems. In each step of BML, one new mixture component is estimated according to the functional gradient of an objective function to ensure that it is added along the direction that maximizes the objective function. Several techniques have been proposed to extend BML from simple mixture models like the Gaussian mixture model (GMM) to the Gaussian mixture hidden Markov model (HMM), including Viterbi approximation for state segmentation, weight decay and sampling boosting to initialize sample weights to avoid overfitting, combination between partial updating and global updating to refine model parameters in each BML iteration, and use of the Bayesian Information Criterion (BIC) for parsimonious modeling. Experimental results on two large-vocabulary continuous speech recognition tasks, namely the WSJ-5k and Switchboard tasks, have shown that the proposed BML yields significant performance gain over the conventional training procedure, especially for small model sizes.

In this paper, we present a novel approach to relax the constraint of stereo-data which is needed in a series of algorithms for noise-robust speech recognition. As a demonstration in SPLICE algorithm, we generate the pseudo-clean features to replace the ideal clean features from one of the stereo channels, by using HMM-based speech synthesis. Experimental results on aurora2 database show that the performance of our approach is comparable with that of SPLICE. Further improvements are achieved by concatenating a bias adaptation algorithm to handle unknown environments. Relative word error rate reductions of 66% and 24% are achieved over the baseline systems in the clean-training and multi-training conditions, respectively.

In this paper, we propose a novel boosted mixture learning (BML) framework for Gaussian mixture HMMs in speech recognition. BML is an incremental method to learn mixture models for classification problem. In each step of BML, one new mixture component is calculated according to functional gradient of an objective function to ensure that it is added along the direction to maximize the objective function the most. Several techniques have been proposed to extend BML from simple mixture models like Gaussian mixture model (GMM) to Gaussian mixture hidden Markov model (HMM), including Viterbi approximation to obtain state segmentation, weight decay to initialize sample weights to avoid overfitting, combining partial updating with global updating of parameters and using Bayesian information criterion (BIC) for parsimonious modeling. Experimental results on the WSJ0 task have shown that the proposed BML yields relative word and sentence error rate reduction of 10.9% and 12.9%, respectively, over the conventional training procedure.

This paper presents a new feature compensation approach to noisy speech recognition by using piecewise linear approximation (PLA) of an explicit model of environmental distortions. Two traditional approaches, namely vector Taylor series (VTS) and MAX approximations, are two special cases of our proposed approach. Formulations for maximum likelihood (ML) estimation of noise model parameters and minimum mean square error (MMSE) estimation of clean speech are derived. A hybrid approach of using different approximations for different types of noisy speech segments is also proposed. Experimental results on Aurora2 and Aurora3 databases demonstrate that the proposed approaches achieve consistently significant improvements in recognition accuracy compared to the traditional VTS-based feature compensation approach.

In this paper, we propose a new feature normalization approach for robust speech recognition. It is found that the shape of speech feature distributions is changed in noisy environments compared with that in the clean condition. So cepstral shape normalization (CSN) which normalizes the shape of feature distributions is performed by exploiting an exponential factor. This method has been proven effective in noisy environments, especially under low SNRs. Experimental results show that the proposed method yields relative word error rate reductions of 38% and 25% on aurora2 and aurora3 databases, respectively, in comparing with those of the conventional mean and variance normalization (MVN). It is also shown CSN consistently outperforms other traditional methods, such as histogram equalization (HEQ) and higher order cepstral moment normalization (HOCMN).

This paper presents a speech enhancement approach derived by using a piecewise linear approximation (PLA) of an explicit model of environmental distortions. PLA is a generalization of two traditional approaches, namely vector Taylor series (VTS) and MAX approximations. Formulations are described for both maximum likelihood (ML) estimation of noise model parameters and minimum mean-squared error (MMSE) estimation of clean speech. Evaluation experiments are conducted to enhance speech signals corrupted by several types of additive noises. Compared to the traditional MAX-approximation based approach, our PLA-based speech enhancement approach achieves better performance in terms of two objective quality measures, namely segmental SNR and log-spectral distortion.

This paper presents a new feature compensation approach to noisy speech recognition by using high-order vector Taylor series (HOVTS) approximation of an explicit model of environmental distortions. Formulations for maximum-likelihood (ML) estimation of both additive noises and convolutional distortions, and minimum mean squared error (MMSE) estimation of clean speech are derived. Experimental results on Aurora2 and Aurora4 benchmark databases, where the modeling assumption of the distortion model is more accurate, demonstrate that the standard HOVTS-based feature compensation approaches achieve consistently significant improvement in recognition accuracy compared to traditional standard first-order VTS-based approach. For a real-world in-vehicle connected digits recognition task on Aurora3 benchmark database where the modeling assumption of the distortion model is less accurate, modifications are necessary to make VTS-based feature compensation approaches work. In this case, the second-order VTS-based approach performs only slightly better than the first-order VTS-based approach.

We propose to use minimum divergence, where acoustic similarity between HMMs is characterized by Kullback-Leibler divergence, for discriminative training. The MD objective function is defined as a posterior weighted divergence measured over the whole training set. Different from our earlier work, where KLD-based acoustic similarity is pre-computed for all initial models and stays invariant in the optimization procedure, here we propose to jointly optimize the whole variable MD by adjusting HMM parameters since MD is a function of the adjusted HMM parameters. An EBW optimization method is derived to minimize the whole MD objective function. The new MD formulation is evaluated on the TIDIGITS and Switchboard databases. Experimental results show that the new MD yields relative word error rate reductions of 62.1% on TIDIGITS and 8.8% on Switchboard databases when compared with the best ML-trained systems. It is also shown the new MD consistently outperforms other discriminative training criteria, such as MPE.

We propose to use Minimum Divergence(MD) as a new measure of errors in discriminative training. To focus on improving discrimination between any two given acoustic models, we refine the error definition in terms of Kullback-Leibler Divergence (KLD) between them. The new measure can be regarded as a modified version of Minimum Phone Error (MPE) but with a higher resolution than just a symbol matching based criterion. Experimental recognition results show the new MD based training yields relative word error rate reductions of 57.8% and 6.1% on TIDigits and Switchboard databases, respectively, in comparing with the ML trained baseline systems. The recognition performance of MD is also shown to be consistently better than that of MPE.