Query rewriting is a crucial task for improving retrieval, especially in professional domains such as law and medicine, where user queries are often underspecified and ambiguous. While large language models (LLMs) offer strong understanding and generation capabilities, existing LLM-based approaches reduce the task to text transformation or expansion, neglecting reasoning to disambiguate queries, which fails to bridge the cognitive gap between user queries and specialized documents. In this paper, we propose Think-Then-Rewrite (TTR), a reinforcement learning based framework that unleashes LLMs' reasoning ability for domain-specific query rewriting. TTR introduces a contrastive mutual information reward to encourage the LLM to generate reasoning processes that effectively distinguish confusing distractors. To boost early-stage training, TTR also constructs golden query rewrites as off‑policy data, providing strong guidance for RL learning. A mixed-policy optimization then combines on-policy and off-policy signals, ensuring both effectiveness and stability. Extensive experiments on legal and medical retrieval benchmarks demonstrate that TTR achieves state-of-the-art performance.

Zhengxi Lu, Fei Tang, Guangyi Liu, Jin Ma, Kaitao Song, Xu Tan, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yifan Yang, Bing Han, Hui Wang, Wei Wang, Ziyang Ma, Long Zhou, Zengrui Jin, Guanrou Yang, Tianrui Wang, Xu Tan, Xie Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent talking avatar generation models have made strides in achieving realistic and accurate lip synchronization with the audio, but often fall short in controlling and conveying detailed expressions and emotions of the avatar, making the generated video less vivid and controllable. In this paper, we propose a text-guided approach for generating emotionally expressive 2D avatars, offering fine-grained control, improved interactivity, and generalizability to the resulting video. Our framework, named InstructAvatar, leverages a natural language interface to control the emotion as well as the facial motion of avatars. Technically, we utilize GPT-4V to design an automatic annotation pipeline, constructing an instruction-video paired training dataset. This is combined with a novel two-branch diffusion-based generator to predict avatars using both audio and text instructions simultaneously. Experimental results demonstrate that InstructAvatar produces results that align well with both conditions, and outperforms existing methods in fine-grained emotion control, lip-sync quality, and naturalness.

Recent advancements in audio generation have been significantly propelled by the capabilities of Large Language Models (LLMs). The existing research on audio LLM has primarily focused on enhancing the architecture and scale of audio language models, as well as leveraging larger datasets, and generally, acoustic codecs, such as EnCodec, are used for audio tokenization. However, these codecs were originally designed for audio compression, which may lead to suboptimal performance in the context of audio LLM. Our research aims to address the shortcomings of current audio LLM codecs, particularly their challenges in maintaining semantic integrity in generated audio. For instance, existing methods like VALL-E, which condition acoustic token generation on text transcriptions, often suffer from content inaccuracies and elevated word error rates (WER) due to semantic misinterpretations of acoustic tokens, resulting in word skipping and errors. To overcome these issues, we propose a straightforward yet effective approach called X-Codec. X-Codec incorporates semantic features from a pre-trained semantic encoder before the Residual Vector Quantization (RVQ) stage and introduces a semantic reconstruction loss after RVQ. By enhancing the semantic ability of the codec, X-Codec significantly reduces WER in speech synthesis tasks and extends these benefits to non-speech applications, including music and sound generation. Our experiments in text-to-speech, music continuation, and text-to-sound tasks demonstrate that integrating semantic information substantially improves the overall performance of language models in audio generation.

In this work, we systematically study music generation conditioned solely on the video. First, we present a large-scale dataset by collecting 360K video-music pairs, including various genres such as movie trailers, advertisements, and documentaries. Furthermore, we propose VidMuse, a simple framework for generating music aligned with video inputs. VidMuse stands out by producing high-fidelity music that is both acoustically and semantically aligned with the video. By incorporating local and global visual cues, VidMuse enables the creation of coherent music tracks that consistently match the video content through Long-Short-Term modeling. Through extensive experiments, VidMuse outperforms existing models in terms of audio quality, diversity, and audio-visual alignment. The code and datasets are available at https://vidmuse.github.io/

Style voice conversion aims to transform the speaking style of source speech into a desired style while keeping the original speaker’s identity. However, previous style voice conversion approaches primarily focus on well-defined domains such as emotional aspects, limiting their practical applications. In this study, we present ZSVC, a novel Zero-shot Style Voice Conversion approach that utilizes a speech codec and a latent diffusion model with speech prompting mechanism to facilitate in-context learning for speaking style conversion. To disentangle speaking style and speaker timbre, we introduce information bottleneck to filter speaking style in the source speech and employ Uncertainty Modeling Adaptive Instance Normalization (UMAdaIN) to perturb the speaker timbre in the style prompt. Moreover, we propose a novel adversarial training strategy to enhance in-context learning and improve style similarity. Experiments conducted on 44,000 hours of speech data demonstrate the superior performance of ZSVC in generating speech with diverse speaking styles in zero-shot scenarios.

In this paper, we explore the application of Large Language Models (LLMs) to the pre-training of music. While the prevalent use of MIDI in music modeling is well-established, our findings suggest that LLMs are inherently more compatible with ABC Notation, which aligns more closely with their design and strengths, thereby enhancing the model's performance in musical composition. To address the challenges associated with misaligned measures from different tracks during generation, we propose the development of a $\underline{S}$ynchronized $\underline{M}$ulti-$\underline{T}$rack ABC Notation ($\textbf{SMT-ABC Notation}$), which aims to preserve coherence across multiple musical tracks. Our contributions include a series of models capable of handling up to 8192 tokens, covering 90\% of the symbolic music data in our training set. Furthermore, we explore the implications of the $\underline{S}$ymbolic $\underline{M}$usic $\underline{S}$caling Law ($\textbf{SMS Law}$) on model performance. The results indicate a promising research direction in music generation, offering extensive resources for further research through our open-source contributions.

Recent advancements in audio language models have underscored the pivotal role of audio tokenization, which converts audio signals into discrete tokens, thereby facilitating the application of language model architectures to the audio domain. In this study, we introduce ALMTokenizer, a novel low-bitrate and semantically rich audio codec tokenizer for audio language models. Prior methods, such as Encodec, typically encode individual audio frames into discrete tokens without considering the use of context information across frames. Unlike these methods, we introduce a novel query-based compression strategy to capture holistic information with a set of learnable query tokens by explicitly modeling the context information across frames. This design not only enables the codec model to capture more semantic information but also encodes the audio signal with fewer token sequences. Additionally, to enhance the semantic information in audio codec models, we introduce the following: (1) A masked autoencoder (MAE) loss, (2) Vector quantization based on semantic priors, and (3) An autoregressive (AR) prediction loss. As a result, ALMTokenizer achieves competitive reconstruction performance relative to state-of-the-art approaches while operating at a lower bitrate. Within the same audio language model framework, ALMTokenizer outperforms previous tokenizers in audio understanding and generation tasks.[https://dongchaoyang.top/ALMTokenizer/]

Symbolic music generation aims to create musical notes, which can help users compose music, such as generating target instrument tracks based on provided source tracks. In practical scenarios where there’s a predefined ensemble of tracks and various composition needs, an efficient and effective generative model that can generate any target tracks based on the other tracks becomes crucial. However, previous efforts have fallen short in addressing this necessity due to limitations in their music representations and models. In this paper, we introduce a framework known as GETMusic, with ``GET'' standing for ``GEnerate music Tracks.'' This framework encompasses a novel music representation ``GETScore'' and a diffusion model ``GETDiff.'' GETScore represents musical notes as tokens and organizes tokens in a 2D structure, with tracks stacked vertically and progressing horizontally over time. At a training step, each track of a music piece is randomly selected as either the target or source. The training involves two processes: In the forward process, target tracks are corrupted by masking their tokens, while source tracks remain as the ground truth; in the denoising process, GETDiff is trained to predict the masked target tokens conditioning on the source tracks. Our proposed representation, coupled with the non-autoregressive generative model, empowers GETMusic to generate music with any arbitrary source-target track combinations.Our experiments demonstrate that the versatile GETMusic outperforms prior works proposed for certain specific composition tasks.

Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, Deqing Yang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Kan Ren, Dongsheng Li, Deqing Yang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Recent advances in text-to-speech synthesis have achieved notable success in generating high-quality short utterances for individual speakers. However, these systems still face challenges when extending their capabilities to long, multi-speaker, and spontaneous dialogues, typical of real-world scenarios such as podcasts. These limitations arise from two primary challenges: 1) long speech: podcasts typically span several minutes, exceeding the upper limit of most existing work; 2) spontaneity: podcasts are marked by their spontaneous, oral nature, which sharply contrasts with formal, written contexts; existing works often fall short in capturing this spontaneity. In this paper, we propose MoonCast, a solution for high-quality zero-shot podcast generation, aiming to synthesize spontaneous podcast-style speech from text-only sources (e.g., stories, technical reports, news in TXT, PDF, or Web URL formats) using the voices of unseen speakers. To enable long audio generation, we employ a language model with parameter, data, and context scaling to process sequences in an innovative format designed for modeling entire multi-speaker, multi-turn speech interactions. To enhance spontaneity, we observe that ASR transcripts capture spontaneous speech details (e.g., filler words indicating hesitations, and specific punctuation and spaces reflecting breathing pauses), suggesting that these transcripts can serve as a partial indicator of speech spontaneity. Building upon this assumption, we utilize a script generation module to generate scripts incorporating these spontaneous elements. Experiments show MoonCast outperforms baselines, with notable improvements in contextual coherence and spontaneity.

In this paper, we propose a novel learning paradigm, termed *Chain-of-Model* (CoM), which incorporates the causal relationship into the hidden states of each layer as a chain style. thereby introducing great scaling efficiency in model training and inference flexibility in deployment.We introduce the concept of *Chain-of-Representation* (CoR), which formulates the hidden states at each layer as a combination of multiple sub-representations (i.e., chains). In each layer, each chain from the output representations can only view all of its preceding chains in the input representations. Consequently, the model built upon CoM framework can progressively scale up the model size by increasing the chains based on the previous models (i.e., chains), and offer multiple sub-models at varying sizes for elastic inference by using different chain numbers. Based on this principle, we devise *Chain-of-Language-Model* (CoLM), which incorporates the idea of CoM into each layer of Transformer architecture. Based on CoLM, we further introduce CoLM-Air by introducing a *KV sharing* mechanism, that computes all keys and values within the first chain and then shares across all chains. This design demonstrates additional extensibility, such as enabling seamless LM switching, prefilling acceleration and so on. Experimental results demonstrate our CoLM family can achieve comparable performance to the standard Transformer, while simultaneously enabling greater flexiblity, such as progressive scaling to improve training efficiency and offer multiple varying model sizes for elastic inference, paving a a new way toward building language models.

Machine learning methods for conditional data generation usually build a mapping from source conditional data X to target data Y. The target Y (e.g., text, speech, music, image, video) is usually high-dimensional and complex, and contains information that does not exist in source data, which hinders effective and efficient learning on the source-target mapping. In this paper, we present a learning paradigm called regeneration learning for data generation, which first generates Y' (an abstraction/representation of Y) from X and then generates Y from Y'. During training, Y' is obtained from Y through either handcrafted rules or self-supervised learning and is used to learn X-->Y' and Y'-->Y. Regeneration learning extends the concept of representation learning to data generation tasks, and can be regarded as a counterpart of traditional representation learning, since 1) regeneration learning handles the abstraction (Y') of the target data Y for data generation while traditional representation learning handles the abstraction (X') of source data X for data understanding; 2) both the processes of Y'-->Y in regeneration learning and X-->X' in representation learning can be learned in a self-supervised way (e.g., pre-training); 3) both the mappings from X to Y' in regeneration learning and from X' to Y in representation learning are simpler than the direct mapping from X to Y. We show that regeneration learning can be a widely-used paradigm for data generation (e.g., text generation, speech recognition, speech synthesis, music composition, image generation, and video generation) and can provide valuable insights into developing data generation methods.

Large Language Models (LLMs) excel in various tasks, but they rely on carefully crafted prompts that often demand substantial human effort. To automate this process, in this paper, we propose a novel framework for discrete prompt optimization, called EvoPrompt, which borrows the idea of evolutionary algorithms (EAs) as they exhibit good performance and fast convergence. To enable EAs to work on discrete prompts, which are natural language expressions that need to be coherent and human-readable, we connect LLMs with EAs. This approach allows us to simultaneously leverage the powerful language processing capabilities of LLMs and the efficient optimization performance of EAs. Specifically, abstaining from any gradients or parameters, EvoPrompt starts from a population of prompts and iteratively generates new prompts with LLMs based on the evolutionary operators, improving the population based on the development set. We optimize prompts for both closed- and open-source LLMs including GPT-3.5 and Alpaca, on 31 datasets covering language understanding, generation tasks, as well as BIG-Bench Hard (BBH) tasks. EvoPrompt significantly outperforms human-engineered prompts and existing methods for automatic prompt generation (e.g., up to 25% on BBH). Furthermore, EvoPrompt demonstrates that connecting LLMs with EAs creates synergies, which could inspire further research on the combination of LLMs and conventional algorithms.

Zero-shot talking avatar generation aims at synthesizing natural talking videos from speech and a single portrait image. Previous methods have relied on domain-specific heuristics such as warping-based motion representation and 3D Morphable Models, which limit the naturalness and diversity of the generated avatars. In this work, we introduce GAIA (Generative AI for Avatar), which eliminates the domain priors in talking avatar generation. In light of the observation that the speech only drives the motion of the avatar while the appearance of the avatar and the background typically remain the same throughout the entire video, we divide our approach into two stages: 1) disentangling each frame into motion and appearance representations; 2) generating motion sequences conditioned on the speech and reference portrait image. We collect a large-scale high-quality talking avatar dataset and train the model on it with different scales (up to 2B parameters). Experimental results verify the superiority, scalability, and flexibility of GAIA as 1) the resulting model beats previous baseline models in terms of naturalness, diversity, lip-sync quality, and visual quality; 2) the framework is scalable since larger models yield better results; 3) it is general and enables different applications like controllable talking avatar generation and text-instructed avatar generation.

Speech conveys more information than text, as the same word can be uttered in various voices to convey diverse information. Compared to traditional text-to-speech (TTS) methods relying on speech prompts (reference speech) for voice variability, using text prompts (descriptions) is more user-friendly since speech prompts can be hard to find or may not exist at all. TTS approaches based on the text prompt face two main challenges: 1) the one-to-many problem, where not all details about voice variability can be described in the text prompt, and 2) the limited availability of text prompt datasets, where vendors and large cost of data labeling are required to write text prompts for speech. In this work, we introduce PromptTTS 2 to address these challenges with a variation network to provide variability information of voice not captured by text prompts, and a prompt generation pipeline to utilize the large language models (LLM) to compose high quality text prompts. Specifically, the variation network predicts the representation extracted from the reference speech (which contains full information about voice variability) based on the text prompt representation. For the prompt generation pipeline, it generates text prompts for speech with a speech language understanding model to recognize voice attributes (e.g., gender, speed) from speech and a large language model to formulate text prompts based on the recognition results. Experiments on a large-scale (44K hours) speech dataset demonstrate that compared to the previous works, PromptTTS 2 generates voices more consistent with text prompts and supports the sampling of diverse voice variability, thereby offering users more choices on voice generation. Additionally, the prompt generation pipeline produces high-quality text prompts, eliminating the large labeling cost. The demo page of PromptTTS 2 is available (https://speechresearch.github.io/prompttts2).

Scaling text-to-speech (TTS) to large-scale, multi-speaker, and in-the-wild datasets is important to capture the diversity in human speech such as speaker identities, prosodies, and styles (e.g., singing). Current large TTS systems usually quantize speech into discrete tokens and use language models to generate these tokens one by one, which suffer from unstable prosody, word skipping/repeating issue, and poor voice quality. In this paper, we develop NaturalSpeech 2, a TTS system that leverages a neural audio codec with residual vector quantizers to get the quantized latent vectors and uses a diffusion model to generate these latent vectors conditioned on text input. To enhance the zero-shot capability that is important to achieve diverse speech synthesis, we design a speech prompting mechanism to facilitate in-context learning in the diffusion model and the duration/pitch predictor. We scale NaturalSpeech 2 to large-scale datasets with 44K hours of speech and singing data and evaluate its voice quality on unseen speakers. NaturalSpeech 2 outperforms previous TTS systems by a large margin in terms of prosody/timbre similarity, robustness, and voice quality in a zero-shot setting, and performs novel zero-shot singing synthesis with only a speech prompt. Audio samples are available at https://naturalspeech2.github.io/.

While recent large-scale text-to-speech (TTS) models have achieved significant progress, they still fall shorts in speech quality, similarity, and prosody. Considering that speech intricately encompasses various attributes (e.g., content, prosody, timbre, and acoustic details) that pose significant challenges for generation, a natural idea is to factorize speech into individual subspaces representing different attributes and generate them individually. Motivated by it, we propose a TTS system with novel factorized diffusion models to generate natural speech in a zero-shot way. Specifically, 1) we design a neural codec with factorized vector quantization (FVQ) to disentangle speech waveform into subspaces of content, prosody, timbre, and acoustic details; 2) we propose a factorized diffusion model, which generates attributes in each subspace following its corresponding prompt. With this factorization design, our method can effectively and efficiently model the intricate speech with disentangled subspaces in a divide-and-conquer way. Experimental results show that our method outperforms the state-of-the-art TTS systems on quality, similarity, prosody, and intelligibility.

Audio generation is a major branch of generative AI research. Compared with prior works in this area that are commonly task-specific with heavy domain knowledge, this paper advocates building universal audio generation models that can handle various tasks in a unified manner. As recent research on large language models (LLMs) has demonstrated their strong ability to handle multiple tasks, this work presents UniAudio, an LLM-based audio generation model that supports a wide range of audio generation tasks. Based on various input conditions, such as phoneme, text description, or audio itself, UniAudio can generate speech, sound, music, and singing voice. The proposed UniAudio is built with 100k hours of multi-source open-available audio data and is scaled to 1B parameters. The audio tokenization method and language model architecture are also specifically designed for both performance and efficiency. Experimentally, UniAuido supports 11 audio generation tasks and achieves competitive results on all tasks consistently. We also show that UniAudio can support new tasks seamlessly via simple fine-tuning.

The latest Text-to-Speech (TTS) systems can produce speech with voice quality and naturalness comparable to human speech. Yet the demand for large amount of high-quality data from target speakers remains a significant challenge. Particularly for long-form expressive reading, target speaker's training speech that covers rich contextual information are needed. In this paper a novel design of context-aware speech pre-trained model is developed for expressive TTS based on contrastive learning. The model can be trained with abundant speech data without explicitly labelled speaker identities. It captures the intricate relationship between the speech expression of a spoken sentence and the contextual text information. By incorporating cross-modal text and speech features into the TTS model, it enables the generation of coherent and expressive speech, which is especially beneficial when there is a scarcity of target speaker data. The pre-trained model is evaluated first in the task of Context-Speech retrieval and then as the integral part of a zero-shot TTS system. Experimental results demonstrate that the pretraining framework effectively learns Context-Speech representations and significantly enhances the expressiveness of synthesized speech. Audio demos are available at: https://ccsp2024.github.io/demo/.

Recent progress in large-scale zero-shot speech synthesis has been significantly advanced by language models and diffusion models. However, the generation process of both methods is slow and computationally intensive. Efficient speech synthesis using a lower computing budget to achieve quality on par with previous work remains a significant challenge. In this paper, we present FlashSpeech, a large-scale zero-shot speech synthesis system with approximately 5% of the inference time compared with previous work. FlashSpeech is built on the latent consistency model and applies a novel adversarial consistency training approach that can train from scratch without the need for a pre-trained diffusion model as the teacher. Furthermore, a new prosody generator module enhances the diversity of prosody, making the rhythm of the speech sound more natural. The generation processes of FlashSpeech can be achieved efficiently with one or two sampling steps while maintaining high audio quality and high similarity to the audio prompt for zero-shot speech generation. Our experimental results demonstrate the superior performance of FlashSpeech. Notably, FlashSpeech can be about 20 times faster than other zero-shot speech synthesis systems while maintaining comparable performance in terms of voice quality and similarity. Furthermore, FlashSpeech demonstrates its versatility by efficiently performing tasks like voice conversion, speech editing, and diverse speech sampling. Audio samples can be found in https://flashspeech.github.io/

Understanding the speaking style, such as the emotion of the interlocutor's speech, and responding with speech in an appropriate style is a natural occurrence in human conversations. However, technically, existing research on speech synthesis and speaking style captioning typically proceeds independently. In this work, an innovative framework, referred to as UniStyle, is proposed to incorporate both the capabilities of speaking style captioning and style-controllable speech synthesizing. Specifically, UniStyle consists of a UniConnector and a style prompt-based speech generator. The role of the UniConnector is to bridge the gap between different modalities, namely speech audio and text descriptions. It enables the generation of text descriptions with speech as input and the creation of style representations from text descriptions for speech synthesis with the speech generator. Besides, to overcome the issue of data scarcity, we propose a two-stage and semi-supervised training strategy, which reduces data requirements while boosting performance. Extensive experiments conducted on open-source corpora demonstrate that UniStyle achieves state-of-the-art performance in speaking style captioning and synthesizes expressive speech with various speaker timbres and speaking styles in a zero-shot manner.

Zhujin Gao, Junliang Guo, Xu Tan, Yongxin Zhu, Fang Zhang, Jiang Bian, Linli Xu. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

In recent years, the remarkable progress of large language models (LLMs) has sparked interest in task automation, which involves decomposing complex tasks described by user instructions into sub-tasks and invoking external tools to execute them, playing a central role in autonomous agents. However, there is a lack of systematic and standardized benchmarks to promote the development of LLMs in task automation. To address this, we introduce TaskBench, a comprehensive framework to evaluate the capability of LLMs in task automation. Specifically, task automation can be divided into three critical stages: task decomposition, tool selection, and parameter prediction. To tackle the complexities inherent in these stages, we introduce the concept of Tool Graph to represent decomposed tasks and adopt a back-instruct method to generate high-quality user instructions. We propose TaskEval, a multi-faceted evaluation methodology that assesses LLM performance across these three stages. Our approach combines automated construction with rigorous human verification, ensuring high consistency with human evaluation. Experimental results demonstrate that TaskBench effectively reflects the capabilities of various LLMs in task automation. It provides insights into model performance across different task complexities and domains, pushing the boundaries of what current models can achieve. TaskBench offers a scalable, adaptable, and reliable benchmark for advancing LLM-based autonomous agents.

Residual networks, as discrete approximations of Ordinary Differential Equations (ODEs), have inspired significant advancements in neural network design, including multistep methods, high-order methods, and multi-particle dynamical systems. The precision of the solution to ODEs significantly affects parameter optimization, thereby impacting model performance. In this work, we present a series of advanced explorations of Transformer architecture design to minimize the error compared to the true ``solution.'' First, we introduce a predictor-corrector learning framework to minimize truncation errors, which consists of a high-order predictor and a multistep corrector. Second, we propose an exponential moving average-based coefficient learning method to strengthen our higher-order predictor. Extensive experiments on large-scale machine translation, abstractive summarization, language modeling, and natural language understanding benchmarks demonstrate the superiority of our approach. On the WMT'14 English-German and English-French tasks, our model achieved BLEU scores of 30.95 and 44.27, respectively. Furthermore, on the OPUS multilingual machine translation task, our model surpasses a robust 3.8B DeepNet by an average of 2.9 SacreBLEU, using only 1/3 parameters. Notably, it also beats LLama models by 5.7 accuracy points on the LM Harness Evaluation.

Continual Pre-Training (CPT) on Large Language Models (LLMs) has been widely used to expand the model’s fundamental understanding of specific downstream domains (e.g., math and code). For the CPT on domain-specific LLMs, one important question is how to choose the optimal mixture ratio between the general-corpus (e.g., Dolma, Slim-pajama) and the downstream domain-corpus. Existing methods usually adopt laborious human efforts by grid-searching on a set of mixture ratios, which require high GPU training consumption costs. Besides, we cannot guarantee the selected ratio is optimal for the specific domain. To address the limitations of existing methods, inspired by the Scaling Law for performance prediction, we propose to investigate the Scaling Law of the Domain-specific Continual Pre-Training (D-CPT Law) to decide the optimal mixture ratio with acceptable training costs for LLMs of different sizes. Specifically, by fitting the D-CPT Law, we can easily predict the general and downstream performance of arbitrary mixture ratios, model sizes, and dataset sizes using small-scale training costs on limited experiments. Moreover, we also extend our standard D-CPT Law on cross-domain settings and propose the Cross-Domain D-CPT Law to predict the D-CPT law of target domains, where very small training costs (about 1\% of the normal training costs) are needed for the target domains. Comprehensive experimental results on six downstream domains demonstrate the effectiveness and generalizability of our proposed D-CPT Law and Cross-Domain D-CPT Law.

Large Language models (LLMs) have demonstrated supreme capabilities in textual understanding and generation, but cannot be directly applied to cross-modal tasks without fine-tuning. This paper proposes a cross-modal in-context learning approach, empowering the frozen LLMs to achieve multiple audio tasks in a few-shot style without any parameter update. Specifically, we propose a novel LLM-driven audio codec model, LLM-Codec, which transfers the audio modality into textual space by representing audio tokens with words or sub-words from the LLM vocabulary, while maintaining high audio reconstruction quality. The key idea is to reduce the modality heterogeneity between text and audio by compressing the audio modality into the well-trained textual space of LLMs. Thus, the audio representation can be viewed as a new \textit{foreign language}, and LLMs can learn the new \textit{foreign language} with several demonstrations. In experiments, we investigate the performance of the proposed approach across multiple audio understanding and generation tasks, \textit{e.g.} speech emotion classification, audio classification, text-to-speech generation, speech enhancement, etc. Experimental results show that LLMs equipped with the LLM-Codec, named as UniAudio 1.5, prompted by only a few examples, can perform effectively in simple scenarios, validating our cross-modal in-context learning approach. To facilitate research on few-shot audio task learning and multi-modal LLMs, we have open-sourced the LLM-Codec model.

This paper introduces Embarrassingly Easy Text-to-Speech (E2 TTS), a fully non-autoregressive zero-shot text-to-speech system that offers human-level naturalness and state-of-the-art speaker similarity and intelligibility. In the E2 TTS framework, the text input is converted into a character sequence with filler tokens. The flow-matching-based mel spectrogram generator is then trained based on the audio infilling task. Unlike many previous works, it does not require additional components (e.g., duration model, grapheme-to-phoneme) or complex techniques (e.g., monotonic alignment search). Despite its simplicity, E2 TTS achieves state-of-the-art zero-shot TTS capabilities that are comparable to or surpass previous works, including Voicebox and NaturalSpeech 3. The simplicity of E2 TTS also allows for flexibility in the input representation. We propose several variants of E2 TTS to improve usability during inference. See https://aka.ms/e2tts/for demo samples.

Neural audio codec models are becoming increasingly important as they serve as tokenizers for audio, enabling efficient transmission or facilitating speech language modeling. The ideal neural audio codec should maintain content, paralinguistics, speaker characteristics, and audio information even at low bitrates. Recently, numerous advanced neural codec models have been proposed. However, codec models are often tested under varying experimental conditions. As a result, we introduce the Codec-SUPERB challenge at SLT 20241, designed to facilitate fair and lightweight comparisons among existing codec models and inspire advancements in the field. This challenge brings together representative speech applications and objective metrics, and carefully selects license-free datasets, sampling them into small sets to reduce evaluation computation costs. This paper presents the challenge’s rules, datasets, participant systems, results, and findings.1https://codecsuperb.github.io/

Text-to-speech (TTS) has made rapid progress in both academia and industry in recent years. Some questions naturally arise that whether a TTS system can achieve human-level quality, how to define/judge that quality, and how to achieve it. In this paper, we answer these questions by first defining the human-level quality based on the statistical significance of subjective measure and introducing appropriate guidelines to judge it, and then developing a TTS system called NaturalSpeech that achieves human-level quality on benchmark datasets. Specifically, we leverage a variational auto-encoder (VAE) for end-to-end text-to-waveform generation, with several key modules to enhance the capacity of the prior from text and reduce the complexity of the posterior from speech, including phoneme pre-training, differentiable duration modeling, bidirectional prior/posterior modeling, and a memory mechanism in VAE. Experimental evaluations on the popular LJSpeech dataset show that our proposed NaturalSpeech achieves -0.01 CMOS (comparative mean opinion score) to human recordings at the sentence level, with Wilcoxon signed rank test at p-level p >> 0.05, which demonstrates no statistically significant difference from human recordings for the first time.

Talking face generation aims at generating photo-realistic video portraits of a target person driven by input audio. According to the nature of audio to lip motions mapping, the same speech content may have different appearances even for the same person at different occasions. Such one-to-many mapping problem brings ambiguity during training and thus causes inferior visual results. Although this one-to-many mapping could be alleviated in part by a two-stage framework (i.e., an audio-to-expression model followed by a neural-rendering model), it is still insufficient since the prediction is produced without enough information (e.g., emotions, wrinkles, etc.). In this paper, we propose MemFace to complement the missing information with an implicit memory and an explicit memory that follow the sense of the two stages respectively. More specifically, the implicit memory is employed in the audio-to-expression model to capture high-level semantics in the audio-expression shared space, while the explicit memory is employed in the neural-rendering model to help synthesize pixel-level details. Our experimental results show that our proposed MemFace surpasses all the state-of-the-art results across multiple scenarios consistently and significantly.

Error correction in automatic speech recognition (ASR) aims to correct those incorrect words in sentences generated by ASR models. Since recent ASR models usually have low word error rate (WER), to avoid affecting originally correct tokens, error correction models should only modify incorrect words, and therefore detecting incorrect words is important for error correction. Previous works on error correction either implicitly detect error words through target-source attention or CTC (connectionist temporal classification) loss, or explicitly locate specific deletion/substitution/insertion errors. However, implicit error detection does not provide clear signal about which tokens are incorrect and explicit error detection suffers from low detection accuracy. In this paper, we propose SoftCorrect with a soft error detection mechanism to avoid the limitations of both explicit and implicit error detection. Specifically, we first detect whether a token is correct or not through a probability produced by a dedicatedly designed language model, and then design a constrained CTC loss that only duplicates the detected incorrect tokens to let the decoder focus on the correction of error tokens. Compared with implicit error detection with CTC loss, SoftCorrect provides explicit signal about which words are incorrect and thus does not need to duplicate every token but only incorrect tokens; compared with explicit error detection, SoftCorrect does not detect specific deletion/substitution/insertion errors but just leaves it to CTC loss. Experiments on AISHELL-1 and Aidatatang datasets show that SoftCorrect achieves 26.1% and 9.4% CER reduction respectively, outperforming previous works by a large margin, while still enjoying fast speed of parallel generation.

Video dubbing aims to translate the original speech in a film or television program into the speech in a target language, which can be achieved with a cascaded system consisting of speech recognition, machine translation and speech synthesis. To ensure the translated speech to be well aligned with the corresponding video, the length/duration of the translated speech should be as close as possible to that of the original speech, which requires strict length control. Previous works usually control the number of words or characters generated by the machine translation model to be similar to the source sentence, without considering the isochronicity of speech as the speech duration of words/characters in different languages varies. In this paper, we propose VideoDubber, a machine translation system tailored for the task of video dubbing, which directly considers the speech duration of each token in translation, to match the length of source and target speech. Specifically, we control the speech length of generated sentence by guiding the prediction of each word with the duration information, including the speech duration of itself as well as how much duration is left for the remaining words. We design experiments on four language directions (German -> English, Spanish -> English, Chinese English), and the results show that VideoDubber achieves better length control ability on the generated speech than baseline methods. To make up the lack of real-world datasets, we also construct a real-world test set collected from films to provide comprehensive evaluations on the video dubbing task.

In this paper, we propose DIFFUSIONNER, which formulates the named entity recognition task as a boundary-denoising diffusion process and thus generates named entities from noisy spans.During training, DIFFUSIONNER gradually adds noises to the golden entity boundaries by a fixed forward diffusion process and learns a reverse diffusion process to recover the entity boundaries.In inference, DIFFU-SIONNER first randomly samples some noisy spans from a standard Gaussian distribution and then generates the named entities by denoising them with the learned reverse diffusion process.The proposed boundary-denoising diffusion process allows progressive refinement and dynamic sampling of entities, empowering DIFFUSIONNER with efficient and flexible entity generation capability.Experiments on multiple flat and nested NER datasets demonstrate that DIFFUSIONNER achieves comparable or even better performance than previous state-of-the-art models 1 .

Dialogue summarization aims to condense the lengthy dialogue into a concise summary, and has recently achieved significant progress.However, the result of existing methods is still far from satisfactory.Previous works indicated that omission is a major factor in affecting the quality of summarization, but few of them have further explored the omission problem, such as how omission affects summarization results and how to detect omission, which is critical for reducing omission and improving summarization quality.Moreover, analyzing and detecting omission relies on summarization datasets with omission labels (i.e., which dialogue utterances are omitted in the summarization), which are not available in the current literature.In this paper, we propose the OLDS dataset, which provides high-quality Omission Labels for Dialogue Summarization.By analyzing this dataset, we find that a large improvement in summarization quality can be achieved by providing ground-truth omission labels for the summarization model to recover omission information, which demonstrates the importance of omission detection for omission mitigation in dialogue summarization.Therefore, we formulate an omission detection task and demonstrate our proposed dataset can support the training and evaluation of this task well.We also call for research action on omission detection based on our proposed datasets.Our dataset and codes are publicly available 1 .

Audio-driven talking face has attracted broad interest from academia and industry recently. However, data acquisition and labeling in audio-driven talking face are labor-intensive and costly. The lack of data resource results in poor synthesis effect. To alleviate this issue, we propose to use TTS (Text-To-Speech) for data augmentation to improve few-shot ability of the talking face system. The misalignment problem brought by the TTS audio is solved with the introduction of soft-DTW, which is first adopted in the talking face task. Moreover, features extracted by HuBERT are explored to utilize underlying information of audio, and found to be superior over other features. The proposed method achieves 17%, 14%, 38% dominance on MSE score, DTW score and user study preference repectively over the baseline model, which shows the effectiveness of improving few-shot learning for talking face system with TTS augmentation.

Using a text description as prompt to guide the generation of text or images (e.g., GPT-3 or DALLE-2) has drawn wide attention recently. Beyond text and image generation, in this work, we explore the possibility of utilizing text descriptions to guide speech synthesis. Thus, we develop a text-to-speech (TTS) system (dubbed as PromptTTS) that takes a prompt with both style and content descriptions as input to synthesize the corresponding speech. Specifically, PromptTTS consists of a style encoder and a content encoder to extract the corresponding representations from the prompt, and a speech decoder to synthesize speech according to the extracted style and content representations. Compared with previous works in controllable TTS that require users to have acoustic knowledge to understand style factors such as prosody and pitch, PromptTTS is more user-friendly since text descriptions are a more natural way to express speech style (e.g., "A lady whispers to her friend slowly"). Given that there is no TTS dataset with prompts, to benchmark the task of PromptTTS, we construct and release a dataset containing prompts with style and content information and the corresponding speech. Experiments show that PromptTTS can generate speech with precise style control and high speech quality. Audio samples and our dataset are publicly available1.

3D Morphable Models (3DMMs) demonstrate great potential for reconstructing faithful and animatable 3D facial surfaces from a single image. The facial surface is influenced by the coarse shape, as well as the static detail (e.g., person-specific appearance) and dynamic detail (e.g., expression-driven wrinkles). Previous work struggles to decouple the static and dynamic details through image-level supervision, leading to reconstructions that are not realistic. In this paper, we aim at high-fidelity 3D face reconstruction and propose HiFace to explicitly model the static and dynamic details. Specifically, the static detail is modeled as the linear combination of a displacement basis, while the dynamic detail is modeled as the linear interpolation of two displacement maps with polarized expressions. We exploit several loss functions to jointly learn the coarse shape and fine details with both synthetic and real-world datasets, which enable HiFace to reconstruct high-fidelity 3D shapes with animatable details. Extensive quantitative and qualitative experiments demonstrate that HiFace presents state-of-the-art reconstruction quality and faithfully recovers both the static and dynamic details. Our project page: https://project-hiface.github.io.

Developing digital sound synthesizers is crucial to the music industry as it provides a low-cost way to produce high-quality sounds with rich timbres. Existing traditional synthesizers often require substantial expertise to determine the overall framework of a synthesizer and the parameters of submodules. Since expert knowledge is hard to acquire, it hinders the flexibility to quickly design and tune digital synthesizers for diverse sounds. In this paper, we propose ``NAS-FM'', which adopts neural architecture search (NAS) to build a differentiable frequency modulation (FM) synthesizer. Tunable synthesizers with interpretable controls can be developed automatically from sounds without any prior expert knowledge and manual operating costs. In detail, we train a supernet with a specifically designed search space, including predicting the envelopes of carriers and modulators with different frequency ratios. An evolutionary search algorithm with adaptive oscillator size is then developed to find the optimal relationship between oscillators and the frequency ratio of FM. Extensive experiments on recordings of different instrument sounds show that our algorithm can build a synthesizer fully automatically, achieving better results than handcrafted synthesizers. Audio samples are available at https://nas-fm.github.io/

While state-of-the-art Text-to-Speech systems can generate natural speech of very high quality at sentence level, they still meet great challenges in speech generation for paragraph / long-form reading.Such deficiencies are due to i) ignorance of cross-sentence contextual information, and ii) high computation and memory cost for long-form synthesis.To address these issues, this work develops a lightweight yet effective TTS system, ContextSpeech.Specifically, we first design a memory-cached recurrence mechanism to incorporate global text and speech context into sentence encoding.Then we construct hierarchically-structured textual semantics to broaden the scope for global context enhancement.Additionally, we integrate linearized self-attention to improve model efficiency.Experiments show that ContextSpeech significantly improves the voice quality and prosody expressiveness in paragraph reading with competitive model efficiency.

While recent research has made significant progress in speech-driven talking face generation, the quality of the generated video still lags behind that of real recordings. One reason for this is the use of handcrafted intermediate representations like facial landmarks and 3DMM coefficients, which are designed based on human knowledge and are insufficient to precisely describe facial movements. Additionally, these methods require an external pretrained model for extracting these representations, whose performance sets an upper bound on talking face generation. To address these limitations, we propose a novel method called DAE-Talker that leverages data-driven latent representations obtained from a diffusion autoencoder (DAE). DAE contains an image encoder that encodes an image into a latent vector and a DDIM-based image decoder that reconstructs the image from it. We train our DAE on talking face video frames and then extract their latent representations as the training target for a Conformer-based speech2latent model. During inference, DAE-Talker first predicts the latents from speech and then generates the video frames with the image decoder in DAE from the predicted latents. This allows DAE-Talker to synthesize full video frames and produce natural head movements that align with the content of speech, rather than relying on a predetermined head pose from a template video. We also introduce pose modelling in speech2latent for pose controllability. Additionally, we propose a novel method for generating continuous video frames with the DDIM-based image decoder trained on individual frames, eliminating the need for modelling the joint distribution of consecutive frames directly. Our experiments show that DAE-Talker outperforms existing popular methods in lip-sync, video fidelity, and pose naturalness. We also conduct ablation studies to analyze the effectiveness of the proposed techniques and demonstrate the pose controllability of DAE-Talker.

With the development of deep learning, multimedia data generation (e.g., image generation, audio synthesis, music composition, and video generation) has attracted a lot of attention. Deep learning methods for data generation usually build a mapping from source condition X to target data Y. The target Y (e.g., image, speech, music, video) is usually high-dimensional and complex, and contains rich information not exist in source data, which hinders the effective and efficient learning on the source-target mapping. Representation learning has achieved rapid progress in the past decade, which is beneficial for data understanding tasks. However, traditional representation learning cannot address the challenges faced by multimedia data generation tasks. This tutorial revisits the learning paradigms for data generation and introduces a paradigm called regeneration learning that can improve the effectiveness and efficiency of multimedia data generation. We show that a variety of tasks in multimedia data generation (e.g., image generation, speech synthesis, music composition, video generation) can benefit from this regeneration learning paradigm, and a lot of recent popular data generation models (e.g., DALL-E 1/2, Stable Diffusion, AudioLM, NaturalSpeech 2, MusicLM) can be covered by this learning paradigm.

Talking-head video editing aims to efficiently insert, delete, and substitute the word of a pre-recorded video through a text transcript editor. The key challenge for this task is obtaining an editing model that generates new talking-head video clips which simultaneously have accurate lip synchronization and motion smoothness. Previous approaches, including 3DMM-based (3D Morphable Model) methods and NeRF-based (Neural Radiance Field) methods, are sub-optimal in that they either require minutes of source videos and days of training time or lack the disentangled control of verbal (e.g., lip motion) and non-verbal (e.g., head pose and expression) representations for video clip insertion. In this work, we fully utilize the video context to design a novel framework for talking-head video editing, which achieves efficiency, disentangled motion control, and sequential smoothness. Specifically, we decompose this framework to motion prediction and motion-conditioned rendering: (1) We first design an animation prediction module that efficiently obtains smooth and lip-sync motion sequences conditioned on the driven speech. This module adopts a non-autoregressive network to obtain context prior and improve the prediction efficiency, and it learns a speech-animation mapping prior with better generalization to novel speech from a multi-identity video dataset. (2) We then introduce a neural rendering module to synthesize the photo-realistic and full-head video frames given the predicted motion sequence. This module adopts a pre-trained head topology and uses only few frames for efficient fine-tuning to obtain a person-specific rendering model. Extensive experiments demonstrate that our method efficiently achieves smoother editing results with higher image quality and lip accuracy using less data than previous methods.

Denoising diffusion probabilistic models (DDPMs) have shown promising performance for speech synthesis. However, a large number of iterative steps are required to achieve high sample quality, which restricts the inference speed. Maintaining sample quality while increasing sampling speed has become a challenging task. In this paper, we propose a Consistency Model-based Speech synthesis method, CoMoSpeech, which achieve speech synthesis through a single diffusion sampling step while achieving high audio quality. The consistency constraint is applied to distill a consistency model from a well-designed diffusion-based teacher model, which ultimately yields superior performances in the distilled CoMoSpeech. Our experiments show that by generating audio recordings by a single sampling step, the CoMoSpeech achieves an inference speed more than 150 times faster than real-time on a single NVIDIA A100 GPU, which is comparable to FastSpeech2, making diffusion-sampling based speech synthesis truly practical. Meanwhile, objective and subjective evaluations on text-to-speech and singing voice synthesis show that the proposed teacher models yield the best audio quality, and the one-step sampling based CoMoSpeech achieves the best inference speed with better or comparable audio quality to other conventional multi-step diffusion model baselines. Audio samples and codes are available at https://comospeech.github. https://comospeech.github.io/.

Solving complicated AI tasks with different domains and modalities is a key step toward artificial general intelligence. While there are numerous AI models available for various domains and modalities, they cannot handle complicated AI tasks autonomously. Considering large language models (LLMs) have exhibited exceptional abilities in language understanding, generation, interaction, and reasoning, we advocate that LLMs could act as a controller to manage existing AI models to solve complicated AI tasks, with language serving as a generic interface to empower this. Based on this philosophy, we present HuggingGPT, an LLM-powered agent that leverages LLMs (e.g., ChatGPT) to connect various AI models in machine learning communities (e.g., Hugging Face) to solve AI tasks. Specifically, we use ChatGPT to conduct task planning when receiving a user request, select models according to their function descriptions available in Hugging Face, execute each subtask with the selected AI model, and summarize the response according to the execution results. By leveraging the strong language capability of ChatGPT and abundant AI models in Hugging Face, HuggingGPT can tackle a wide range of sophisticated AI tasks spanning different modalities and domains and achieve impressive results in language, vision, speech, and other challenging tasks, which paves a new way towards the realization of artificial general intelligence.

Audio editing is applicable for various purposes, such as adding background sound effects, replacing a musical instrument, and repairing damaged audio. Recently, some diffusion-based methods achieved zero-shot audio editing by using a diffusion and denoising process conditioned on the text description of the output audio. However, these methods still have some problems: 1) they have not been trained on editing tasks and cannot ensure good editing effects; 2) they can erroneously modify audio segments that do not require editing; 3) they need a complete description of the output audio, which is not always available or necessary in practical scenarios. In this work, we propose AUDIT, an instruction-guided audio editing model based on latent diffusion models. Specifically, \textbf{AUDIT} has three main design features: 1) we construct triplet training data (instruction, input audio, output audio) for different audio editing tasks and train a diffusion model using instruction and input (to be edited) audio as conditions and generating output (edited) audio; 2) it can automatically learn to only modify segments that need to be edited by comparing the difference between the input and output audio; 3) it only needs edit instructions instead of full target audio descriptions as text input. AUDIT achieves state-of-the-art results in both objective and subjective metrics for several audio editing tasks (e.g., adding, dropping, replacement, inpainting, super-resolution). Demo samples are available at https://audit-demopage.github.io/.

Data in the real world tends to exhibit a long-tailed label distribution, which poses great challenges for the training of neural networks in visual recognition. Existing methods tackle this problem mainly from the perspective of data quantity, i.e., the number of samples in each class. To be specific, they pay more attention to tail classes, like applying larger adjustments to the logit. However, in the training process, the quantity and difficulty of data are two intertwined and equally crucial problems. For some tail classes, the features of their instances are distinct and discriminative, which can also bring satisfactory accuracy; for some head classes, although with sufficient samples, the high semantic similarity with other classes and lack of discriminative features will bring bad accuracy. Based on these observations, we propose Adaptive Logit Adjustment Loss (ALA Loss) to apply an adaptive adjusting term to the logit. The adaptive adjusting term is composed of two complementary factors: 1) quantity factor, which pays more attention to tail classes, and 2) difficulty factor, which adaptively pays more attention to hard instances in the training process. The difficulty factor can alleviate the over-optimization on tail yet easy instances and under-optimization on head yet hard instances. The synergy of the two factors can not only advance the performance on tail classes even further, but also promote the accuracy on head classes. Unlike previous logit adjusting methods that only concerned about data quantity, ALA Loss tackles the long-tailed problem from a more comprehensive, fine-grained and adaptive perspective. Extensive experimental results show that our method achieves the state-of-the-art performance on challenging recognition benchmarks, including ImageNet-LT, iNaturalist 2018, and Places-LT.

Non-autoregressive text to speech (NAR-TTS) models have attracted much attention from both academia and industry due to their fast generation speed.One limitation of NAR-TTS models is that they ignore the correlation in time and frequency domains while generating speech mel-spectrograms, and thus cause blurry and over-smoothed results.In this work, we revisit this over-smoothing problem from a novel perspective: the degree of over-smoothness is determined by the gap between the complexity of data distributions and the capability of modeling methods.Both simplifying data distributions and improving modeling methods can alleviate the problem.Accordingly, we first study methods reducing the complexity of data distributions.Then we conduct a comprehensive study on NAR-TTS models that use some advanced modeling methods.Based on these studies, we find that 1) methods that provide additional condition inputs reduce the complexity of data distributions to model, thus alleviating the over-smoothing problem and achieving better voice quality.2) Among advanced modeling methods, Laplacian mixture loss performs well at modeling multimodal distributions and enjoys its simplicity, while GAN and Glow achieve the best voice quality while suffering from increased training or model complexity.3) The two categories of methods can be combined to further alleviate the over-smoothness and improve the voice quality.4) Our experiments on the multi-speaker dataset lead to similar conclusions as above and providing more variance information can reduce the difficulty of modeling the target data distribution and alleviate the requirements for model capacity.

Typical generative dialogue models utilize the dialogue history to generate the response.However, since one dialogue utterance can often be appropriately answered by multiple distinct responses, generating a desired response solely based on the historical information is not easy.Intuitively, if the chatbot can foresee in advance what the user would talk about (i.e., the dialogue future) after receiving its response, it could possibly provide a more informative response.Accordingly, we propose a novel dialogue generation framework named ProphetChat that utilizes the simulated dialogue futures in the inference phase to enhance response generation.To enable the chatbot to foresee the dialogue future, we design a beam-search-like roll-out strategy for dialogue future simulation using a typical dialogue generation model and a dialogue selector.With the simulated futures, we then utilize the ensemble of a history-to-response generator and a future-to-response generator to jointly generate a more informative response.Experiments on two popular open-domain dialogue datasets demonstrate that ProphetChat can generate better responses over strong baselines, which validates the advantages of incorporating the simulated dialogue futures.

Zeqian Ju, Peiling Lu, Xu Tan, Rui Wang, Chen Zhang, Songruoyao Wu, Kejun Zhang, Xiang-Yang Li, Tao Qin, Tie-Yan Liu. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

Text error correction aims to correct the errors in text sequences such as those typed by humans or generated by speech recognition models.Previous error correction methods usually take the source (incorrect) sentence as encoder input and generate the target (correct) sentence through the decoder.Since the error rate of the incorrect sentence is usually low (e.g., 10%), the correction model can only learn to correct on limited error tokens but trivially copy on most tokens (correct tokens), which harms the effective training of error correction.In this paper, we argue that the correct tokens should be better utilized to facilitate effective training and then propose a simple yet effective masking strategy to achieve this goal.Specifically, we randomly mask out a part of the correct tokens in the source sentence and let the model learn to not only correct the original error tokens but also predict the masked tokens based on their context information.Our method enjoys several advantages: 1) it alleviates trivial copy; 2) it leverages effective training signals from correct tokens; 3) it is a plug-and-play module and can be applied to different models and tasks.Experiments on spelling error correction and speech recognition error correction on Mandarin datasets and grammar error correction on English datasets with both autoregressive and non-autoregressive generation models show that our method improves the correction accuracy consistently 1 .

Denoising diffusion probabilistic models (diffusion models for short) require a large number of iterations in inference to achieve the generation quality that matches or surpasses the state-of-the-art generative models, which invariably results in slow inference speed. Previous approaches aim to optimize the choice of inference schedule over a few iterations to speed up inference. However, this results in reduced generation quality, mainly because the inference process is optimized separately, without jointly optimizing with the training process. In this paper, we propose InferGrad, a diffusion model for vocoder that incorporates inference process into training, to reduce the inference iterations while maintaining high generation quality. More specifically, during training, we generate data from random noise through a reverse process under inference schedules with a few iterations, and impose a loss to minimize the gap between the generated and ground-truth data samples. Then, unlike existing approaches, the training of InferGrad considers the inference process. The advantages of InferGrad are demonstrated through experiments on the LJSpeech dataset showing that InferGrad achieves better voice quality than the baseline WaveGrad under same conditions while maintaining the same voice quality as the baseline but with 3x speedup (2 iterations for InferGrad vs 6 iterations for WaveGrad).

We propose a novel robust and efficient Speech-to-Animation (S2A) approach for synchronized facial animation generation in human-computer interaction. Compared with conventional approaches, the proposed approach utilizes phonetic posteriorgrams (PPGs) of spoken phonemes as input to ensure the cross-language and cross-speaker ability, and introduces corresponding prosody features (i.e. pitch and energy) to further enhance the expression of generated animation. Mixture-of-experts (MOE)-based Transformer is employed to better model contextual information while provide significant optimization on computation efficiency. Experiments demonstrate the effectiveness of the proposed approach on both objective and subjective evaluation with 17× inference speedup compared with the state-of-the-art approach.

In the development of neural text-to-speech systems, model pre-training with a large amount of non-target speakers’ data is a common approach. However, in terms of ultimately achieved system performance for target speaker(s), the actual benefits of model pre-training are uncertain and unstable, depending very much on the quantity and text content of training data. This study aims to understand better why and how model pre-training can positively contribute to TTS system performance. It is postulated that the pre-training process plays a critical role in learning text-related variation in speech, while further training with the target speaker’s data aims to capture the speaker-related variation. Different test sets are created with varying degrees of similarity to target speaker data in terms of text content. Experiments show that leveraging a speaker-independent TTS trained on speech data with diverse text content can improve the target speaker TTS on domain-mismatched text. We also attempt to reduce the amount of pre-training data for a new text domain and improve the data and computational efficiency. It is found that the TTS system could achieve comparable performance when the pre-training data is reduced to 1/8 of its original size.

Weight sharing is a popular approach to reduce the training cost of neural architecture search (NAS) by reusing the weights of shared operators from previously trained child models. However, the rank correlation between the estimated accuracy and ground truth accuracy of those child models is low due to the interference among different child models caused by weight sharing. In this paper, we investigate the interference issue by sampling different child models and calculating the gradient similarity of shared operators, and observe that: 1) the interference on a shared operator between two child models is positively correlated with the number of different operators between them; 2) the interference is smaller when the inputs and outputs of the shared operator are more similar. Inspired by these two observations, we propose two approaches to mitigate the interference: 1) rather than randomly sampling child models for optimization, we propose a gradual modification scheme by modifying one operator between adjacent optimization steps to minimize the interference on the shared operators; 2) forcing the inputs and outputs of the operator across all child models to be similar to reduce the interference. Experiments on a BERT search space verify that mitigating interference via each of our proposed methods improves the rank correlation of super-pet and combining both methods can achieve better results. Our discovered architecture outperforms RoBERTa$_{\rm base}$ by 1.1 and 0.6 points and ELECTRA$_{\rm base}$ by 1.6 and 1.1 points on the dev and test set of GLUE benchmark. Extensive results on the BERT compression, reading comprehension and large-scale image classification tasks also demonstrate the effectiveness and generality of our proposed methods.

Adaptive text to speech (TTS) can synthesize new voices in zero-shot scenarios efficiently, by using a well-trained source TTS model without adapting it on the speech data of new speakers. Considering seen and unseen speakers have diverse characteristics, zero-shot adaptive TTS requires strong generalization ability on speaker characteristics, which brings modeling challenges. In this paper, we develop AdaSpeech 4, a zero-shot adaptive TTS system for high-quality speech synthesis. We model the speaker characteristics systematically to improve the generalization on new speakers. Generally, the modeling of speaker characteristics can be categorized into three steps: extracting speaker representation, taking this speaker representation as condition, and synthesizing speech/mel-spectrogram given this speaker representation. Accordingly, we improve the modeling in three steps: 1) To extract speaker representation with better generalization, we factorize the speaker characteristics into basis vectors and extract speaker representation by weighted combining of these basis vectors through attention. 2) We leverage conditional layer normalization to integrate the extracted speaker representation to TTS model. 3) We propose a novel supervision loss based on the distribution of basis vectors to maintain the corresponding speaker characteristics in generated mel-spectrograms. Without any fine-tuning, AdaSpeech 4 achieves better voice quality and similarity than baselines in multiple datasets.

Lyric-to-melody generation, which generates melody according to given lyrics, is one of the most important automatic music composition tasks. With the rapid development of deep learning, previous works address this task with end-to-end neural network models. However, deep learning models cannot well capture the strict but subtle relationships between lyrics and melodies, which compromises the harmony between lyrics and generated melodies. In this paper, we propose ReLyMe, a method that incorporates Relationships between Lyrics and Melodies from music theory to ensure the harmony between lyrics and melodies. Specifically, we first introduce several principles that lyrics and melodies should follow in terms of tone, rhythm, and structure relationships. These principles are then integrated into neural network lyric-to-melody models by adding corresponding constraints during the decoding process to improve the harmony between lyrics and melodies. We use a series of objective and subjective metrics to evaluate the generated melodies. Experiments on both English and Chinese song datasets show the effectiveness of ReLyMe, demonstrating the superiority of incorporating lyric-melody relationships from the music domain into neural lyric-to-melody generation.

Kexun Zhang, Rui Wang, Xu Tan, Junliang Guo, Yi Ren, Tao Qin, Tie-Yan Liu. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Binaural audio plays a significant role in constructing immersive augmented and virtual realities. As it is expensive to record binaural audio from the real world, synthesizing them from mono audio has attracted increasing attention. This synthesis process involves not only the basic physical warping of the mono audio, but also room reverberations and head/ear related filtration, which, however, are difficult to accurately simulate in traditional digital signal processing. In this paper, we formulate the synthesis process from a different perspective by decomposing the binaural audio into a common part that shared by the left and right channels as well as a specific part that differs in each channel. Accordingly, we propose BinauralGrad, a novel two-stage framework equipped with diffusion models to synthesize them respectively. Specifically, in the first stage, the common information of the binaural audio is generated with a single-channel diffusion model conditioned on the mono audio, based on which the binaural audio is generated by a two-channel diffusion model in the second stage. Combining this novel perspective of two-stage synthesis with advanced generative models (i.e., the diffusion models), the proposed BinauralGrad is able to generate accurate and high-fidelity binaural audio samples. Experiment results show that on a benchmark dataset, BinauralGrad outperforms the existing baselines by a large margin in terms of both object and subject evaluation metrics (Wave L2: $0.128$ vs. $0.157$, MOS: $3.80$ vs. $3.61$). The generated audio samples\footnote{\url{https://speechresearch.github.io/binauralgrad}} and code\footnote{\url{https://github.com/microsoft/NeuralSpeech/tree/master/BinauralGrad}} are available online.

Sentence scoring aims at measuring the likelihood score of a sentence and is widely used in many natural language processing scenarios, like reranking, which is to select the best sentence from multiple candidates. Previous works on sentence scoring mainly adopted either causal language modeling (CLM) like GPT or masked language modeling (MLM) like BERT, which have some limitations: 1) CLM only utilizes unidirectional information for the probability estimation of a sentence without considering bidirectional context, which affects the scoring quality; 2) MLM can only estimate the probability of partial tokens at a time and thus requires multiple forward passes to estimate the probability of the whole sentence, which incurs large computation and time cost. In this paper, we propose \textit{Transcormer} -- a Transformer model with a novel \textit{sliding language modeling} (SLM) for sentence scoring. Specifically, our SLM adopts a triple-stream self-attention mechanism to estimate the probability of all tokens in a sentence with bidirectional context and only requires a single forward pass. SLM can avoid the limitations of CLM (only unidirectional context) and MLM (multiple forward passes) and inherit their advantages, and thus achieve high effectiveness and efficiency in scoring. Experimental results on multiple tasks demonstrate that our method achieves better performance than other language modelings.

Symbolic music generation aims to generate music scores automatically. A recent trend is to use Transformer or its variants in music generation, which is, however, suboptimal, because the full attention cannot efficiently model the typically long music sequences (e.g., over 10,000 tokens), and the existing models have shortcomings in generating musical repetition structures. In this paper, we propose Museformer, a Transformer with a novel fine- and coarse-grained attention for music generation. Specifically, with the fine-grained attention, a token of a specific bar directly attends to all the tokens of the bars that are most relevant to music structures (e.g., the previous 1st, 2nd, 4th and 8th bars, selected via similarity statistics); with the coarse-grained attention, a token only attends to the summarization of the other bars rather than each token of them so as to reduce the computational cost. The advantages are two-fold. First, it can capture both music structure-related correlations via the fine-grained attention, and other contextual information via the coarse-grained attention. Second, it is efficient and can model over 3X longer music sequences compared to its full-attention counterpart. Both objective and subjective experimental results demonstrate its ability to generate long music sequences with high quality and better structures.

Automatic song writing aims to compose a song (lyric and/or melody) by machine, which is an interesting topic in both academia and industry. In automatic song writing, lyric-to-melody generation and melody-to-lyric generation are two important tasks, both of which usually suffer from the following challenges: 1) the paired lyric and melody data are limited, which affects the generation quality of the two tasks, considering a lot of paired training data are needed due to the weak correlation between lyric and melody; 2) Strict alignments are required between lyric and melody, which relies on specific alignment modeling. In this paper, we propose SongMASS to address the above challenges, which leverages masked sequence to sequence (MASS) pre-training and attention based alignment modeling for lyric-to-melody and melody-to-lyric generation. Specifically, 1) we extend the original sentence-level MASS pre-training to song level to better capture long contextual information in music, and use a separate encoder and decoder for each modality (lyric or melody); 2) we leverage sentence-level attention mask and token-level attention constraint during training to enhance the alignment between lyric and melody. During inference, we use a dynamic programming strategy to obtain the alignment between each word/syllable in lyric and note in melody. We pre-train SongMASS on unpaired lyric and melody datasets, and both objective and subjective evaluations demonstrate that SongMASS generates lyric and melody with significantly better quality than the baseline method.

Existing speech to speech translation systems heavily rely on the text of target language: they usually translate source language either to target text and then synthesize target speech from text, or directly to target speech with target text for auxiliary training. However, those methods cannot be applied to unwritten target languages, which have no written text or phoneme available. In this paper, we develop a translation system for unwritten languages, named as UWSpeech, which converts target unwritten speech into discrete tokens with a converter, and then translates source-language speech into target discrete tokens with a translator, and finally synthesizes target speech from target discrete tokens with an inverter. We propose a method called XL-VAE, which enhances vector quantized variational autoencoder (VQ-VAE) with cross-lingual (XL) speech recognition, to train the converter and inverter of UWSpeech jointly. Experiments on Fisher Spanish-English conversation translation dataset show that UWSpeech outperforms direct translation and VQ-VAE baseline by about 16 and 10 BLEU points respectively, which demonstrate the advantages and potentials of UWSpeech.

Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, Tie-Yan Liu. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Mean opinion score (MOS) is a popular subjective metric to assess the quality of synthesized speech, and usually involves multiple human judges to evaluate each speech utterance. To reduce the labor cost in MOS test, multiple methods have been proposed to automatically predict MOS scores. To our knowledge, for a speech utterance, all previous works only used the average of multiple scores from different judges as the training target and discarded the score of each individual judge, which did not well exploit the precious MOS training data. In this paper, we propose MBNet, a MOS predictor with a mean subnet and a bias subnet to better utilize every judge score in MOS datasets, where the mean subnet is used to predict the mean score of each utterance similar to that in previous works, and the bias subnet to predict the bias score (the difference between the mean score and each individual judge score) and capture the personal preference of individual judges. Experiments show that compared with MOSNet baseline that only leverages mean score for training, MBNet improves the system-level spearmans rank correlation coefficient (SRCC) by 2.9% on VCC 2018 dataset and 6.7% on VCC 2016 dataset.

Text to speech (TTS) has been broadly used to synthesize natural and intelligible speech in different scenarios. Deploying TTS in various end devices such as mobile phones or embedded devices requires extremely small memory usage and inference latency. While non-autoregressive TTS models such as FastSpeech have achieved significantly faster inference speed than autoregressive models, their model size and inference latency are still large for the deployment in resource constrained devices. In this paper, we propose LightSpeech, which leverages neural architecture search (NAS) to automatically design more lightweight and efficient models based on FastSpeech. We first profile the components of current Fast-Speech model and carefully design a novel search space containing various lightweight and potentially effective architectures. Then NAS is utilized to automatically discover well performing architectures within the search space. Experiments show that the model discovered by our method achieves 15x model compression ratio and 6.5x inference speedup on CPU with on par voice quality. Audio demos are provided at https://speechresearch.github.io/lightspeech.

In this paper, we propose MixSpeech, a simple yet effective data augmentation method based on mixup for automatic speech recognition (ASR). MixSpeech trains an ASR model by taking a weighted combination of two different speech features (e.g., mel-spectrograms or MFCC) as the input, and recognizing both text sequences, where the two recognition losses use the same combination weight. We apply MixSpeech on two popular end-to-end speech recognition models including LAS (Listen, Attend and Spell) and Transformer, and conduct experiments on several low-resource datasets including TIMIT, WSJ, and HKUST. Experimental results show that MixSpeech achieves better accuracy than the baseline models without data augmentation, and outperforms a strong data augmentation method SpecAugment on these recognition tasks. Specifically, MixSpeech outperforms SpecAugment with a relative PER improvement of 10.6% on TIMIT dataset, and achieves a strong WER of 4.7% on WSJ dataset.

Text to speech (TTS) is widely used to synthesize personal voice for a target speaker, where a well-trained source TTS model is fine-tuned with few paired adaptation data (speech and its transcripts) on this target speaker. However, in many scenarios, only untranscribed speech data is available for adaptation, which brings challenges to the previous TTS adaptation pipelines (e.g., AdaSpeech). In this paper, we develop AdaSpeech 2, an adaptive TTS system that only leverages untranscribed speech data for adaptation. Specifically, we introduce a mel-spectrogram encoder to a well-trained TTS model to conduct speech reconstruction, and at the same time constrain the output sequence of the mel-spectrogram encoder to be close to that of the original phoneme encoder. In adaptation, we use untranscribed speech data for speech reconstruction and only fine-tune the TTS decoder. AdaSpeech 2 has two advantages: 1) Pluggable: our system can be easily applied to existing trained TTS models without re-training. 2) Effective: our system achieves on-par voice quality with the transcribed TTS adaptation (e.g., AdaSpeech) with the same amount of untranscribed data, and achieves better voice quality than previous untranscribed adaptation methods1.

While neural-based text to speech (TTS) models can synthesize natural and intelligible voice, they usually require high-quality speech data, which is costly to collect. In many scenarios, only noisy speech of a target speaker is available, which presents challenges for TTS model training for this speaker. Previous works usually address the challenge using two methods: 1) training the TTS model using the speech denoised with an enhancement model; 2) taking a single noise embedding as input when training with noisy speech. However, they usually cannot handle speech with real-world complicated noise such as those with high variations along time. In this paper, we develop DenoiSpeech, a TTS system that can synthesize clean speech for a speaker with noisy speech data. In DenoiSpeech, we handle real-world noisy speech by modeling the fine-grained frame-level noise with a noise condition module, which is jointly trained with the TTS model. Experimental results on real-world data show that DenoiSpeech outperforms the previous two methods by 0.31 and 0.66 MOS respectively.1

Neural approaches have achieved state-of-the-art accuracy on machine translation but suffer from the high cost of collecting large scale parallel data. Thus, a lot of research has been conducted for neural machine translation (NMT) with very limited parallel data, i.e., the low-resource setting. In this paper, we provide a survey for low-resource NMT and classify related works into three categories according to the auxiliary data they used: (1) exploiting monolingual data of source and/or target languages, (2) exploiting data from auxiliary languages, and (3) exploiting multi-modal data. We hope that our survey can help researchers to better understand this field and inspire them to design better algorithms, and help industry practitioners to choose appropriate algorithms for their applications.

End-to-end automatic speech recognition (ASR) can achieve promising performance with large-scale training data. However, it is known that domain mismatch between training and testing data often leads to a degradation of recognition accuracy. In this work, we focus on the unsupervised domain adaptation for ASR and propose CMatch, a Character-level distribution matching method to perform fine-grained adaptation between each character in two domains. First, to obtain labels for the features belonging to each character, we achieve frame-level label assignment using the Connectionist Temporal Classification (CTC) pseudo labels. Then, we match the character-level distributions using Maximum Mean Discrepancy. We train our algorithm using the self-training technique. Experiments on the Libri-Adapt dataset show that our proposed approach achieves 14.39% and 16.50% relative Word Error Rate (WER) reduction on both cross-device and cross-environment ASR. We also comprehensively analyze the different strategies for frame-level label assignment and Transformer adaptations.

While pre-trained language models (e.g., BERT) have achieved impressive results on different natural language processing tasks, they have large numbers of parameters and suffer from big computational and memory costs, which make them difficult for real-world deployment. Therefore, model compression is necessary to reduce the computation and memory cost of pre-trained models. In this work, we aim to compress BERT and address the following two challenging practical issues: (1) The compression algorithm should be able to output multiple compressed models with different sizes and latencies, in order to support devices with different memory and latency limitations; (2) The algorithm should be downstream task agnostic, so that the compressed models are generally applicable for different downstream tasks. We leverage techniques in neural architecture search (NAS) and propose NAS-BERT, an efficient method for BERT compression. NAS-BERT trains a big supernet on a carefully designed search space containing a variety of architectures and outputs multiple compressed models with adaptive sizes and latency. Furthermore, the training of NAS-BERT is conducted on standard self-supervised pre-training tasks (e.g., masked language model) and does not depend on specific downstream tasks. Thus, the compressed models can be used across various downstream tasks. The technical challenge of NAS-BERT is that training a big supernet on the pre-training task is extremely costly. We employ several techniques including block-wise search, search space pruning, and performance approximation to improve search efficiency and accuracy. Extensive experiments on GLUE and SQuAD benchmark datasets demonstrate that NAS-BERT can find lightweight models with better accuracy than previous approaches, and can be directly applied to different downstream tasks with adaptive model sizes for different requirements of memory or latency.

AI music composition is one of the most attractive and important topics in artificial intelligence, music, and multimedia. The typical tasks in AI music composition include melody generation, song writing, accompaniment generation, arrangement, performance generation, timbre rendering, sound generation, and singing voice synthesis, which cover different modalities (e.g., symbolic music score, sound) and well match to the theme of ACM Multimedia. As the rapid development of artificial intelligence techniques such as content creation and deep learning, AI based music composition has achieved rapid progress, but still encountered a lot of challenges. A thorough introduction and review on the basics, the research progress, as well as how to address the challenges in AI music composition are timely and necessary for a broad audience working on artificial intelligence, music, and multimedia. In this tutorial, we will first introduce the background of AI music composition, including music basics and deep learning techniques for music composition. Then we will introduce AI music composition from two perspectives: 1) key components, which include music score generation, music performance generation, and music sound generation; 2) advanced topics, which include music structure/form/style/emotion modeling, timbre synthesis/transfer/mixing, etc. At last, we will point out some research challenges and future directions in AI music composition. This tutorial can serve both academic researchers and industry practitioners working on AI music composition.

Neural Transducer (e.g., RNN-T) has been widely used in automatic speech recognition (ASR) due to its capabilities of efficiently modeling monotonic alignments between input and output sequences and naturally supporting streaming inputs. Considering that monotonic alignments are also critical to text to speech (TTS) synthesis and streaming TTS is also an important application scenario, in this work, we explore the possibility of applying Transducer to TTS and more. However, it is challenging because it is difficult to trade off the emission (continuous mel-spectrogram prediction) probability and transition (ASR Transducer predicts blank token to indicate transition to next input) probability when calculating the output probability lattice in Transducer, and it is not easy to learn the alignments between text and speech through the output probability lattice. We propose SpeechTransducer (Speech-T for short), a Transformer based Transducer model that 1) uses a new forward algorithm to separate the transition prediction from the continuous mel-spectrogram prediction when calculating the output probability lattice, and uses a diagonal constraint in the probability lattice to help the alignment learning; 2) supports both full-sentence or streaming TTS by adjusting the look-ahead context; and 3) further supports both TTS and ASR together for the first time, which enjoys several advantages including fewer parameters as well as streaming synthesis and recognition in a single model. Experiments on LJSpeech datasets demonstrate that Speech-T 1) is more robust than the attention based autoregressive TTS model due to its inherent monotonic alignments between text and speech; 2) naturally supports streaming TTS with good voice quality; and 3) enjoys the benefit of joint modeling TTS and ASR in a single network.

Error correction techniques have been used to refine the output sentences from automatic speech recognition (ASR) models and achieve a lower word error rate (WER) than original ASR outputs. Previous works usually use a sequence-to-sequence model to correct an ASR output sentence autoregressively, which causes large latency and cannot be deployed in online ASR services. A straightforward solution to reduce latency, inspired by non-autoregressive (NAR) neural machine translation, is to use an NAR sequence generation model for ASR error correction, which, however, comes at the cost of significantly increased ASR error rate. In this paper, observing distinctive error patterns and correction operations (i.e., insertion, deletion, and substitution) in ASR, we propose FastCorrect, a novel NAR error correction model based on edit alignment. In training, FastCorrect aligns each source token from an ASR output sentence to the target tokens from the corresponding ground-truth sentence based on the edit distance between the source and target sentences, and extracts the number of target tokens corresponding to each source token during edition/correction, which is then used to train a length predictor and to adjust the source tokens to match the length of the target sentence for parallel generation. In inference, the token number predicted by the length predictor is used to adjust the source tokens for target sequence generation. Experiments on the public AISHELL-1 dataset and an internal industrial-scale ASR dataset show the effectiveness of FastCorrect for ASR error correction: 1) it speeds up the inference by 6-9 times and maintains the accuracy (8-14% WER reduction) compared with the autoregressive correction model; and 2) it outperforms the popular NAR models adopted in neural machine translation and text edition by a large margin.

Non-autoregressive translation (NAT) models remove the dependence on previous target tokens and generate all target tokens in parallel, resulting in significant inference speedup but at the cost of inferior translation accuracy compared to autoregressive translation (AT) models. Considering that AT models have higher accuracy and are easier to train than NAT models, and both of them share the same model configurations, a natural idea to improve the accuracy of NAT models is to transfer a well-trained AT model to an NAT model through fine-tuning. However, since AT and NAT models differ greatly in training strategy, straightforward fine-tuning does not work well. In this work, we introduce curriculum learning into fine-tuning for NAT. Specifically, we design a curriculum in the fine-tuning process to progressively switch the training from autoregressive generation to non-autoregressive generation. Experiments on four benchmark translation datasets show that the proposed method achieves good improvement (more than 1 BLEU score) over previous NAT baselines in terms of translation accuracy, and greatly speed up (more than 10 times) the inference process over AT baselines.

In this work, we develop SimulSpeech, an endto-end simultaneous speech to text translation system which translates speech in source language to text in target language concurrently.SimulSpeech consists of a speech encoder, a speech segmenter and a text decoder, where 1) the segmenter builds upon the encoder and leverages a connectionist temporal classification (CTC) loss to split the input streaming speech in real time, 2) the encoder-decoder attention adopts a wait-k strategy for simultaneous translation.SimulSpeech is more challenging than previous cascaded systems (with simultaneous automatic speech recognition (ASR) and simultaneous neural machine translation (NMT)).We introduce two novel knowledge distillation methods to ensure the performance: 1) Attention-level knowledge distillation transfers the knowledge from the multiplication of the attention matrices of simultaneous NMT and ASR models to help the training of the attention mechanism in SimulSpeech; 2) Data-level knowledge distillation transfers the knowledge from the full-sentence NMT model and also reduces the complexity of data distribution to help on the optimization of Simul-Speech.Experiments on MuST-C English-Spanish and English-German spoken language translation datasets show that SimulSpeech achieves reasonable BLEU scores and lower delay compared to full-sentence end-to-end speech to text translation (without simultaneous translation), and better performance than the two-stage cascaded simultaneous translation model in terms of BLEU scores and translation delay.

Non-autoregressive (NAR) models generate all the tokens of a sequence in parallel, resulting in faster generation speed compared to their autoregressive (AR) counterparts but at the cost of lower accuracy.Different techniques including knowledge distillation and source-target alignment have been proposed to bridge the gap between AR and NAR models in various tasks such as neural machine translation (NMT), automatic speech recognition (ASR), and text to speech (TTS).With the help of those techniques, NAR models can catch up with the accuracy of AR models in some tasks but not in some others.In this work, we conduct a study to understand the difficulty of NAR sequence generation and try to answer: (1) Why NAR models can catch up with AR models in some tasks but not all?(2) Why techniques like knowledge distillation and source-target alignment can help NAR models.Since the main difference between AR and NAR models is that NAR models do not use dependency among target tokens while AR models do, intuitively the difficulty of NAR sequence generation heavily depends on the strongness of dependency among target tokens.To quantify such dependency, we propose an analysis model called CoMMA to characterize the difficulty of different NAR sequence generation tasks.We have several interesting findings: 1) Among the NMT, ASR and TTS tasks, ASR has the most target-token dependency while TTS has the least.2) Knowledge distillation reduces the target-token dependency in target sequence and thus improves the accuracy of NAR models.3) Source-target alignment constraint encourages dependency * Equal contribution.

This paper introduces a new end-to-end text-to-speech (E2E-TTS) toolkit named ESPnet-TTS, which is an extension of the open-source speech processing toolkit ESPnet. The toolkit supports state-of- the-art E2E-TTS models, including Tacotron 2, Transformer TTS, and FastSpeech, and also provides recipes inspired by the Kaldi automatic speech recognition (ASR) toolkit. The recipes are based on the design unified with the ESPnet ASR recipe, providing high reproducibility. The toolkit also provides pre-trained models and samples of all of the recipes so that users can use it as a baseline. Furthermore, the unified design enables the integration of ASR functions with TTS, e.g., ASR-based objective evaluation and semi- supervised learning with both ASR and TTS models. This paper describes the design of the toolkit and experimental evaluation in comparison with other toolkits. The experimental results show that our models can achieve state-of-the-art performance comparable to the other latest toolkits, resulting in a mean opinion score (MOS) of 4.25 on the LJSpeech dataset. The toolkit is publicly available at https://github.com/espnet/espnet.

Non-autoregressive translation (NAT) achieves faster inference speed but at the cost of worse accuracy compared with autoregressive translation (AT). Since AT and NAT can share model structure and AT is an easier task than NAT due to the explicit dependency on previous target-side tokens, a natural idea is to gradually shift the model training from the easier AT task to the harder NAT task. To smooth the shift from AT training to NAT training, in this paper, we introduce semi-autoregressive translation (SAT) as intermediate tasks. SAT contains a hyperparameter k, and each k value defines a SAT task with different degrees of parallelism. Specially, SAT covers AT and NAT as its special cases: it reduces to AT when k=1 and to NAT when k=N (N is the length of target sentence). We design curriculum schedules to gradually shift k from 1 to N, with different pacing functions and number of tasks trained at the same time. We called our method as task-level curriculum learning for NAT (TCL-NAT). Experiments on IWSLT14 De-En, IWSLT16 En-De, WMT14 En-De and De-En datasets show that TCL-NAT achieves significant accuracy improvements over previous NAT baselines and reduces the performance gap between NAT and AT models to 1-2 BLEU points, demonstrating the effectiveness of our proposed method.

Neural machine translation (NMT) generates the next target token given as input the previous ground truth target tokens during training while the previous generated target tokens during inference, which causes discrepancy between training and inference as well as error propagation, and affects the translation accuracy. In this paper, we introduce an error correction mechanism into NMT, which corrects the error information in the previous generated tokens to better predict the next token. Specifically, we introduce two-stream self-attention from XLNet into NMT decoder, where the query stream is used to predict the next token, and meanwhile the content stream is used to correct the error information from the previous predicted tokens. We leverage scheduled sampling to simulate the prediction errors during training. Experiments on three IWSLT translation datasets and two WMT translation datasets demonstrate that our method achieves improvements over Transformer baseline and scheduled sampling. Further experimental analyses also verify the effectiveness of our proposed error correction mechanism to improve the translation quality.

Transformer-based text to speech (TTS) model (e.g., Transformer TTS [1], FastSpeech [2]) has shown the advantages of training and inference efficiency over Tacotron [3]) due to its parallel computation in training and/or inference.However, the parallel computation increases the difficulty while learning the alignment between text and speech in Transformer, which is further magnified in the multi-speaker scenario with noisy data and diverse speakers, and hinders the applicability of Transformer for multi-speaker TTS.In this paper, we develop a robust and high-quality multispeaker Transformer TTS system called MultiSpeech, with several specially designed components/techniques to improve textto-speech alignment: 1) a diagonal constraint on the weight matrix of encoder-decoder attention in both training and inference; 2) layer normalization on phoneme embedding in encoder to better preserve position information; 3) a bottleneck in decoder pre-net to prevent copy between consecutive speech frames.Experiments on VCTK and LibriTTS multi-speaker datasets demonstrate the effectiveness of MultiSpeech: 1) it synthesizes more robust and better quality multi-speaker voice than naive Transformer based TTS; 2) with a MutiSpeech model as the teacher, we obtain a strong multi-speaker FastSpeech model with almost zero quality degradation while enjoying extremely fast inference speed.

This paper presents XiaoiceSing, a high-quality singing voice synthesis system which employs an integrated network for spectrum, F0 and duration modeling.We follow the main architecture of FastSpeech while proposing some singing-specific design: 1) Besides phoneme ID and position encoding, features from musical score (e.g.note pitch and length) are also added.2) To attenuate off-key issues, we add a residual connection in F0 prediction.3) In addition to the duration loss of each phoneme, the duration of all the phonemes in a musical note is accumulated to calculate the syllable duration loss for rhythm enhancement.Experiment results show that XiaoiceSing outperforms the baseline system of convolutional neural networks by 1.44 MOS on sound quality, 1.18 on pronunciation accuracy and 1.38 on naturalness respectively.In two A/B tests, the proposed F0 and duration modeling methods achieve 97.3% and 84.3% preference rate over baseline respectively, which demonstrates the overwhelming advantages of XiaoiceSing.

In this paper, we develop DeepSinger, a multi-lingual multi-singer singing voice synthesis (SVS) system, which is built from scratch using singing training data mined from music websites. The pipeline of DeepSinger consists of several steps, including data crawling, singing and accompaniment separation, lyrics-to-singing alignment, data filtration, and singing modeling. Specifically, we design a lyrics-to-singing alignment model to automatically extract the duration of each phoneme in lyrics starting from coarse-grained sentence level to fine-grained phoneme level, and further design a multi-lingual multi-singer singing model based on a feed-forward Transformer to directly generate linear-spectrograms from lyrics, and synthesize voices using Griffn-Lim. DeepSinger has several advantages over previous SVS systems: 1) to the best of our knowledge, it is the first SVS system that directly mines training data from music websites, 2) the lyrics-to-singing alignment model further avoids any human efforts for alignment labeling and greatly reduces labeling cost, 3) the singing model based on a feed-forward Transformer is simple and efficient, by removing the complicated acoustic feature modeling in parametric synthesis and leveraging a reference encoder to capture the timbre of a singer from noisy singing data, and 4) it can synthesize singing voices in multiple languages and multiple singers. We evaluate DeepSinger on our mined singing dataset that consists of about 92 hours data from 89 singers on three languages (Chinese, Cantonese and English). The results demonstrate that with the singing data purely mined from the Web, DeepSinger can synthesize high-quality singing voices in terms of both pitch accuracy and voice naturalness. Our audio samples are shown in https://speechresearch.github.io/deepsinger/.

Speech synthesis (text to speech, TTS) and recognition (automatic speech recognition, ASR) are important speech tasks, and require a large amount of text and speech pairs for model training. However, there are more than 6,000 languages in the world and most languages are lack of speech training data, which poses significant challenges when building TTS and ASR systems for extremely low-resource languages. In this paper, we develop LRSpeech, a TTS and ASR system under the extremely low-resource setting, which can support rare languages with low data cost. LRSpeech consists of three key techniques: 1) pre-training on rich-resource languages and fine-tuning on low-resource languages; 2) dual transformation between TTS and ASR to iteratively boost the accuracy of each other; 3) knowledge distillation to customize the TTS model on a high-quality target-speaker voice and improve the ASR model on multiple voices. We conduct experiments on an experimental language (English) and a truly low-resource language (Lithuanian) to verify the effectiveness of LRSpeech. Experimental results show that LRSpeech 1) achieves high quality for TTS in terms of both intelligibility (more than $98%$ intelligibility rate) and naturalness (above 3.5 mean opinion score (MOS)) of the synthesized speech, which satisfy the requirements for industrial deployment, 2) achieves promising recognition accuracy for ASR, and 3) last but not least, uses extremely low-resource training data. We also conduct comprehensive analyses on LRSpeech with different amounts of data resources, and provide valuable insights and guidances for industrial deployment. We are currently deploying LRSpeech into a commercialized cloud speech service to support TTS on more rare languages.

Lip reading aims to recognize text from talking lip, while lip generation aims to synthesize talking lip according to text, which is a key component in talking face generation and is a dual task of lip reading. Both tasks require a large amount of paired lip video and text training data, and perform poorly in low-resource scenarios with limited paired training data. In this paper, we develop DualLip, a system that jointly improves lip reading and generation by leveraging the task duality and using unlabeled text and lip video data. The key ideas of the DualLip include: 1) Generate lip video from unlabeled text using a lip generation model, and use the pseudo data pairs to improve lip reading; 2) Generate text from unlabeled lip video using a lip reading model, and use the pseudo data pairs to improve lip generation. To leverage the benefit of DualLip on lip generation, we further extend DualLip to talking face generation with two additionally introduced components: lip to face generation and text to speech generation, which share the same duration for synchronization. Experiments on GRID and TCD-TIMIT datasets demonstrate the effectiveness of DualLip on improving lip reading, lip generation and talking face generation by utilizing unlabeled data, especially in low-resource scenarios. Specifically, on the GRID dataset, the lip generation model in our DualLip system trained with only 10% paired data and 90% unpaired data surpasses the performance of that trained with the whole paired data, and our lip reading model achieves 1.16% character error rate and 2.71% word error rate, outperforming the state-of-the-art models using the same amount of paired data.

In pop music, accompaniments are usually played by multiple instruments (tracks) such as drum, bass, string and guitar, and can make a song more expressive and contagious by arranging together with its melody. Previous works usually generate multiple tracks separately and the music notes from different tracks not explicitly depend on each other, which hurts the harmony modeling. To improve harmony, in this paper, we propose a novel MUlti-track MIDI representation (MuMIDI), which enables simultaneous multi-track generation in a single sequence and explicitly models the dependency of the notes from different tracks. While this greatly improves harmony, unfortunately, it enlarges the sequence length and brings the new challenge of long-term music modeling. We further introduce two new techniques to address this challenge: 1) We model multiple note attributes (e.g., pitch, duration, velocity) of a musical note in one step instead of multiple steps, which can shorten the length of a MuMIDI sequence. 2) We introduce extra long-context as memory to capture long-term dependency in music. We call our system for pop music accompaniment generation as PopMAG. We evaluate PopMAG on multiple datasets (LMD, FreeMidi and CPMD, a private dataset of Chinese pop songs) with both subjective and objective metrics. The results demonstrate the effectiveness of PopMAG for multi-track harmony modeling and long-term context modeling. Specifically, PopMAG wins 42%/38%/40% votes when comparing with ground truth musical pieces on LMD, FreeMidi and CPMD datasets respectively and largely outperforms other state-of-the-art music accompaniment generation models and multi-track MIDI representations in terms of subjective and objective metrics.

Neural architecture search (NAS) relies on a good controller to generate better architectures or predict the accuracy of given architectures. However, training the controller requires both abundant and high-quality pairs of architectures and their accuracy, while it is costly to evaluate an architecture and obtain its accuracy. In this paper, we propose SemiNAS, a semi-supervised NAS approach that leverages numerous unlabeled architectures (without evaluation and thus nearly no cost). Specifically, SemiNAS 1) trains an initial accuracy predictor with a small set of architecture-accuracy data pairs; 2) uses the trained accuracy predictor to predict the accuracy of large amount of architectures (without evaluation); and 3) adds the generated data pairs to the original data to further improve the predictor. The trained accuracy predictor can be applied to various NAS algorithms by predicting the accuracy of candidate architectures for them. SemiNAS has two advantages: 1) It reduces the computational cost under the same accuracy guarantee. On NASBench-101 benchmark dataset, it achieves comparable accuracy with gradient-based method while using only 1/7 architecture-accuracy pairs. 2) It achieves higher accuracy under the same computational cost. It achieves 94.02% test accuracy on NASBench-101, outperforming all the baselines when using the same number of architectures. On ImageNet, it achieves 23.5% top-1 error rate (under 600M FLOPS constraint) using 4 GPU-days for search. We further apply it to LJSpeech text to speech task and it achieves 97% intelligibility rate in the low-resource setting and 15% test error rate in the robustness setting, with 9%, 7% improvements over the baseline respectively.

BERT adopts masked language modeling (MLM) for pre-training and is one of the most successful pre-training models. Since BERT neglects dependency among predicted tokens, XLNet introduces permuted language modeling (PLM) for pre-training to address this problem. However, XLNet does not leverage the full position information of a sentence and thus suffers from position discrepancy between pre-training and fine-tuning. In this paper, we propose MPNet, a novel pre-training method that inherits the advantages of BERT and XLNet and avoids their limitations. MPNet leverages the dependency among predicted tokens through permuted language modeling (vs. MLM in BERT), and takes auxiliary position information as input to make the model see a full sentence and thus reducing the position discrepancy (vs. PLM in XLNet). We pre-train MPNet on a large-scale dataset (over 160GB text corpora) and fine-tune on a variety of down-streaming tasks (GLUE, SQuAD, etc). Experimental results show that MPNet outperforms MLM and PLM by a large margin, and achieves better results on these tasks compared with previous state-of-the-art pre-trained methods (e.g., BERT, XLNet, RoBERTa) under the same model setting. We attach the code in the supplemental materials.

Maximum-likelihood estimation (MLE) is widely used in sequence to sequence tasks for model training. It uniformly treats the generation/prediction of each target token as multiclass classification, and yields non-smooth prediction probabilities: in a target sequence, some tokens are predicted with small probabilities while other tokens are with large probabilities. According to our empirical study, we find that the non-smoothness of the probabilities results in low quality of generated sequences. In this paper, we propose a sentence-wise regularization method which aims to output smooth prediction probabilities for all the tokens in the target sequence. Our proposed method can automatically adjust the weights and gradients of each token in one sentence to ensure the predictions in a sequence uniformly well. Experiments on three neural machine translation tasks and one text summarization task show that our method outperforms conventional MLE loss on all these tasks and achieves promising BLEU scores on WMT14 English-German and WMT17 Chinese-English translation task.

Non-autoregressive translation (NAT) models, which remove the dependence on previous target tokens from the inputs of the decoder, achieve significantly inference speedup but at the cost of inferior accuracy compared to autoregressive translation (AT) models. Previous work shows that the quality of the inputs of the decoder is important and largely impacts the model accuracy. In this paper, we propose two methods to enhance the decoder inputs so as to improve NAT models. The first one directly leverages a phrase table generated by conventional SMT approaches to translate source tokens to target tokens, which are then fed into the decoder as inputs. The second one transforms source-side word embeddings to target-side word embeddings through sentence-level alignment and word-level adversary learning, and then feeds the transformed word embeddings into the decoder as inputs. Experimental results show our method largely outperforms the NAT baseline (Gu et al. 2017) by 5.11 BLEU scores on WMT14 English-German task and 4.72 BLEU scores on WMT16 English-Romanian task.

Sharing source and target side vocabularies and word embeddings has been a popular practice in neural machine translation (briefly, NMT) for similar languages (e.g., English to French or German translation). The success of such wordlevel sharing motivates us to move one step further: we consider model-level sharing and tie the whole parts of the encoder and decoder of an NMT model. We share the encoder and decoder of Transformer (Vaswani et al. 2017), the state-of-the-art NMT model, and obtain a compact model named Tied Transformer. Experimental results demonstrate that such a simple method works well for both similar and dissimilar language pairs. We empirically verify our framework for both supervised NMT and unsupervised NMT: we achieve a 35.52 BLEU score on IWSLT 2014 German to English translation, 28.98/29.89 BLEU scores on WMT 2014 English to German translation without/with monolingual data, and a 22.05 BLEU score on WMT 2016 unsupervised German to English translation.

Unsupervised neural machine translation (NMT) has attracted a lot of attention recently.While state-of-the-art methods for unsupervised translation usually perform well between similar languages (e.g., English-German translation), they perform poorly between distant languages, because unsupervised alignment does not work well for distant languages.In this work, we introduce unsupervised pivot translation for distant languages, which translates a language to a distant language through multiple hops, and the unsupervised translation on each hop is relatively easier than the original direct translation.We propose a learning to route (LTR) method to choose the translation path between the source and target languages.LTR is trained on language pairs whose best translation path is available and is applied to the unseen language pairs for path selection.Experiments on 20 languages and 294 distant language pairs demonstrate the advantages of the unsupervised pivot translation for distant languages, as well as the effectiveness of the proposed LTR for path selection.Specifically, in the best case, LTR achieves an improvement of 5.58 BLEU points over the conventional direct unsupervised method.

Polyphone disambiguation aims to select the correct pronunciation for a polyphonic word from several candidates, which is important for text-to-speech synthesis. Since the pronunciation of a polyphonic word is usually decided by its context, polyphone disambiguation can be regarded as a language understanding task. Inspired by the success of BERT for language understanding, we propose to leverage pre-trained BERT models for polyphone disambiguation. However, BERT models are usually too heavy to be served online, in terms of both memory cost and inference speed. In this work, we focus on efficient model for polyphone disambiguation and propose a two-stage knowledge distillation method that transfers the knowledge from a heavy BERT model in both pre-training and fine-tuning stages to a lightweight BERT model, in order to reduce online serving cost. Experiments on Chinese and English polyphone disambiguation datasets demonstrate that our method reduces model parameters by a factor of 5 and improves inference speed by 7 times, while nearly matches the classification accuracy (95.4% on Chinese and 98.1% on English) to the original BERT model.

Xu Tan, Jiale Chen, Di He, Yingce Xia, Tao Qin, Tie-Yan Liu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Text to speech (TTS) and automatic speech recognition (ASR) are two dual tasks in speech processing and both achieve impressive performance thanks to the recent advance in deep learning and large amount of aligned speech and text data. However, the lack of aligned data poses a major practical problem for TTS and ASR on low-resource languages. In this paper, by leveraging the dual nature of the two tasks, we propose an almost unsupervised learning method that only leverages few hundreds of paired data and extra unpaired data for TTS and ASR. Our method consists of the following components: (1) denoising auto-encoder, which reconstructs speech and text sequences respectively to develop the capability of language modeling both in speech and text domain; (2) dual transformation, where the TTS model transforms the text $y$ into speech $\hat{x}$, and the ASR model leverages the transformed pair $(\hat{x},y)$ for training, and vice versa, to boost the accuracy of the two tasks; (3) bidirectional sequence modeling, which address the error propagation problem especially in the long speech and text sequence when training with few paired data; (4) a unified model structure, which combines all the above components for TTS and ASR based on Transformer model. Our method achieves 99.84% in terms of word level intelligible rate and 2.68 MOS for TTS, and 11.7% PER for ASR on LJSpeech dataset, by leveraging only 200 paired speech and text data (about 20 minutes audio), together with extra unpaired speech and text data.

Pre-training and fine-tuning, e.g., BERT \citep{devlin2018bert}, have achieved great success in language understanding by transferring knowledge from rich-resource pre-training task to the low/zero-resource downstream tasks. Inspired by the success of BERT, we propose MAsked Sequence to Sequence pre-training (MASS) for the encoder-decoder based language generation tasks. MASS adopts the encoder-decoder framework to reconstruct a sentence fragment given the remaining part of the sentence: its encoder takes a sentence with randomly masked fragment (several consecutive tokens) as input, and its decoder tries to predict this masked fragment. In this way, MASS can jointly train the encoder and decoder to develop the capability of representation extraction and language modeling. By further fine-tuning on a variety of zero/low-resource language generation tasks, including neural machine translation, text summarization and conversational response generation (3 tasks and totally 8 datasets), MASS achieves significant improvements over the baselines without pre-training or with other pre-training methods. Especially, we achieve the state-of-the-art accuracy (30.02 in terms of BLEU score) on the unsupervised English-French translation, even beating the early attention-based supervised model \citep{bahdanau2015neural}.

Image-to-image translation, which transfers an image from a source domain to a target one, has attracted much attention in both academia and industry. The major approach is to adopt an encoder-decoder based framework, where the encoder extracts features from the input image and then the decoder decodes the features and generates an image in the target domain as the output. In this paper, we go beyond this learning framework by considering an additional polishing step on the output image. Polishing an image is very common in human's daily life, such as editing and beautifying a photo in Photoshop after taking/generating it by a digital camera. Such a deliberation process is shown to be very helpful and important in practice and thus we believe it will also be helpful for image translation. Inspired by the success of deliberation network in natural language processing, we extend deliberation process to the field of image translation. We verify our proposed method on four two-domain translation tasks and one multi-domain translation task. Both the qualitative and quantitative results demonstrate the effectiveness of our method.

Grapheme-to-phoneme (G2P) conversion is an important task in automatic speech recognition and text-to-speech systems.Recently, G2P conversion is viewed as a sequence to sequence task and modeled by RNN or CNN based encoderdecoder framework.However, previous works do not consider the practical issues when deploying G2P model in the production system, such as how to leverage additional unlabeled data to boost the accuracy, as well as reduce model size for online deployment.In this work, we propose token-level ensemble distillation for G2P conversion, which can (1) boost the accuracy by distilling the knowledge from additional unlabeled data, and (2) reduce the model size but maintain the high accuracy, both of which are very practical and helpful in the online production system.We use token-level knowledge distillation, which results in better accuracy than the sequence-level counterpart.What is more, we adopt the Transformer instead of RNN or CNN based models to further boost the accuracy of G2P conversion.Experiments on the publicly available CMU-Dict dataset and an internal English dataset demonstrate the effectiveness of our proposed method.Particularly, our method achieves 19.88% WER on CMUDict dataset, outperforming the previous works by more than 4.22% WER, and setting the new state-of-the-art results.

Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from the mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the problem of word skipping and repeating in particularly hard cases, and can adjust voice speed smoothly. Most importantly, compared with autoregressive Transformer TTS, our model speeds up mel-spectrogram generation by 270x and the end-to-end speech synthesis by 38x. Therefore, we call our model FastSpeech.

Sequence generation tasks, such as neural machine translation (NMT) and abstractive summarization, usually suffer from exposure bias as well as the error propagation problem due to the autoregressive training and generation. Many previous works have discussed the relationship between error propagation and theaccuracy dropproblem (i.e., the right part of the generated sentence is often worse than its left part in left-to-right decoding models). In this paper, taking NMT as a typical sequence generation task, we measure the accuracy of the generated sentence with various metrics and conduct a series of analyses to deeply understand the accuracy drop problem. We obtain several interesting findings. First, The role of error propagation on accuracy drop is overstated in the literature, although it is indeed a cause to the accuracy drop problem. Second, Characteristics of a language play a more important role in causing the accuracy drop problem: the left part of the generated sentence in a right-branching language (e.g., English) is more likely to be more accurate than its right part, while the right part is more accurate for a left-branching language (e.g., Japanese). Our discoveries are also confirmed on other generation tasks (e.g., image captioning, abstractive summarization and language modeling) with multiple left/right-branching languages, as well as in various model structures.

Encoder-decoder based Sequence to Sequence learning (S2S) has made remarkable progress in recent years. Different network architectures have been used in the encoder/decoder. Among them, Convolutional Neural Networks (CNN) and Self Attention Networks (SAN) are the prominent ones. The two architectures achieve similar performances but use very different ways to encode and decode context: CNN use convolutional layers to focus on the local connectivity of the sequence, while SAN uses self-attention layers to focus on global semantics. In this work we propose Double Path Networks for Sequence to Sequence learning (DPN-S2S), which leverage the advantages of both models by using double path information fusion. During the encoding step, we develop a double path architecture to maintain the information coming from different paths with convolutional layers and self-attention layers separately. To effectively use the encoded context, we develop a gated attention fusion module and use it to automatically pick up the information needed during the decoding step, which is also a double path network. By deeply integrating the two paths, both types of information are combined and well exploited. Experiments show that our proposed method can significantly improve the performance of sequence to sequence learning over state-of-the-art systems.

Neural machine translation usually adopts autoregressive models and suffers from exposure bias as well as the consequent error propagation problem.Many previous works have discussed the relationship between error propagation and the accuracy drop (i.e., the left part of the translated sentence is often better than its right part in left-to-right decoding models) problem.In this paper, we conduct a series of analyses to deeply understand this problem and get several interesting findings.(1) The role of error propagation on accuracy drop is overstated in the literature, although it indeed contributes to the accuracy drop problem.(2) Characteristics of a language play a more important role in causing the accuracy drop: the left part of the translation result in a right-branching language (e.g., English) is more likely to be more accurate than its right part, while the right part is more accurate for a left-branching language (e.g., Japanese).Our discoveries are confirmed on different model structures including Transformer and RNN, and in other sequence generation tasks such as text summarization.

Many artificial intelligence tasks appear in dual forms like English$\leftrightarrow$French translation and speech$\leftrightarrow$text transformation. Existing dual learning schemes, which are proposed to solve a pair of such dual tasks, explore how to leverage such dualities from data level. In this work, we propose a new learning framework, model-level dual learning, which takes duality of tasks into consideration while designing the architectures for the primal/dual models, and ties the model parameters that playing similar roles in the two tasks. We study both symmetric and asymmetric model-level dual learning. Our algorithms achieve significant improvements on neural machine translation and sentiment analysis.

As an important and challenging problem in machine learning and computer vision, neural network acceleration essentially aims to enhance the computational efficiency without sacrificing the model accuracy too much. In this paper, we propose a progressive blockwise learning scheme for teacher-student model distillation at the subnetwork block level. The proposed scheme is able to distill the knowledge of the entire teacher network by locally extracting the knowledge of each block in terms of progressive blockwise function approximation. Furthermore, we propose a structure design criterion for the student subnetwork block, which is able to effectively preserve the original receptive field from the teacher network. Experimental results demonstrate the effectiveness of the proposed scheme against the state-of-the-art approaches.

Yanyao Shen, Xu Tan, Di He, Tao Qin, Tie-Yan Liu. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Continuous word representation (aka word embedding) is a basic building block in many neural network-based models used in natural language processing tasks. Although it is widely accepted that words with similar semantics should be close to each other in the embedding space, we find that word embeddings learned in several tasks are biased towards word frequency: the embeddings of high-frequency and low-frequency words lie in different subregions of the embedding space, and the embedding of a rare word and a popular word can be far from each other even if they are semantically similar. This makes learned word embeddings ineffective, especially for rare words, and consequently limits the performance of these neural network models. In order to mitigate the issue, in this paper, we propose a neat, simple yet effective adversarial training method to blur the boundary between the embeddings of high-frequency words and low-frequency words. We conducted comprehensive studies on ten datasets across four natural language processing tasks, including word similarity, language modeling, machine translation and text classification. Results show that we achieve higher performance than the baselines in all tasks.

Neural Machine Translation (NMT) has achieved remarkable progress with the quick evolvement of model structures. In this paper, we propose the concept of layer-wise coordination for NMT, which explicitly coordinates the learning of hidden representations of the encoder and decoder together layer by layer, gradually from low level to high level. Specifically, we design a layer-wise attention and mixed attention mechanism, and further share the parameters of each layer between the encoder and decoder to regularize and coordinate the learning. Experiments show that combined with the state-of-the-art Transformer model, layer-wise coordination achieves improvements on three IWSLT and two WMT translation tasks. More specifically, our method achieves 34.43 and 29.01 BLEU score on WMT16 English-Romanian and WMT14 English-German tasks, outperforming the Transformer baseline.

In this paper, structured visual feature learning aims at exploiting the intrinsic structural properties of mutually correlated multimedia collections (e.g., video frames or facial images) to learn a more effective feature representation for multimedia data classification. We pose structured visual feature learning as a problem of supervised tensor factorization (STF), which is capable of effectively learning multi-view visual features from structural tensorial multimedia data. In mathematics , STF is formulated as a joint optimization framework of probabilistic inference and$\epsilon $-insensitive support vector regression. As a result, the feature representation obtained by STF not only preserves the intrinsic multi-view structural information on tensorial multimedia data, but also includes the discriminative information derived from the max-margin learning process. Using the learned discriminative visual features, we conduct a set of multimedia classification experiments on several challenging datasets, including images and videos, which demonstrate the effectiveness of our method.

Non-negative tensor factorization (NTF) has attracted great attention in the machine learning community. In this paper, we extend traditional non-negative tensor factorization into a supervised discriminative decomposition, referred as Supervised Non-negative Tensor Factorization with Maximum-Margin Constraint(SNTFM2). SNTFM2 formulates the optimal discriminative factorization of non-negative tensorial data as a coupled least-squares optimization problem via a maximum-margin method. As a result, SNTFM2 not only faithfully approximates the tensorial data by additive combinations of the basis, but also obtains a strong generalization power to discriminative analysis (in particularfor classification in this paper). The experimental results show the superiority of our proposed model over state-of-the-art techniques on both toy and real world data sets.