Fakhraddin Alwajih, El Moatez Billah Nagoudi, Gagan Bhatia, Abdelrahman Mohamed, Muhammad Abdul-Mageed. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

We introduce VOICECRAFT, a token infilling neural codec language model, that achieves state-of-the-art performance on both speech editing and zero-shot text-to-speech (TTS) on audiobooks, internet videos, and podcasts 1 .VOICECRAFT employs a Transformer decoder architecture and introduces a token rearrangement procedure that combines causal masking and delayed stacking to enable generation within an existing sequence.On speech editing tasks, VOICECRAFT produces edited speech that is nearly indistinguishable from unedited recordings in terms of naturalness, as evaluated by humans; for zero-shot TTS, our model outperforms prior SotA models including VALL-E and the popular commercial model XTTS v2.Crucially, the models are evaluated on challenging and realistic datasets, that consist of diverse accents, speaking styles, recording conditions, and background noise and music, and our model performs consistently well compared to other models and real recordings.In particular, for speech editing evaluation, we introduce a high quality, challenging, and realistic dataset named REALEDIT.We encourage readers to listen to the demos at https: //jasonppy.github.io/VoiceCraft_web.

Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar El-Shangiti, Hiba Zayed, Mohamedou Cheikh Tourad, Rahaf Alhamouri, Rwaa Assi, Aisha Alraeesi, Hour Mohamed, Fakhraddin Alwajih, Abdelrahman Mohamed, Abdellah El Mekki, El Moatez Billah Nagoudi, Benelhadj Djelloul Mama Saadia, Hamzah A. Alsayadi, Walid Al-Dhabyani, Sara Shatnawi, Yasir Ech-chammakhy, Amal Makouar, Yousra Berrachedi, Mustafa Jarrar, Shady Shehata, Ismail Berrada, Muhammad Abdul-Mageed. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Data-driven unit discovery in self-supervised learning (SSL) of speech has embarked on a new era of spoken language processing. Yet, the discovered units often remain in phonetic space and speech units beyond phonemes are largely underexplored. Here, we demonstrate that a syllabic organization emerges in learning sentence-level representation of speech. In particular, we adopt "self-distillation" objective to fine-tune the pretrained HuBERT with an aggregator token that summarizes the entire sentence. Without any supervision, the resulting model draws definite boundaries in speech, and the representations across frames exhibit salient syllabic structures. We demonstrate that this emergent structure largely corresponds to the ground truth syllables. Furthermore, we propose a new benchmark task, Spoken Speech ABX, for evaluating sentence-level representation of speech. When compared to previous models, our model outperforms in both unsupervised syllable discovery and learning sentence-level representation. Together, we demonstrate that the self-distillation of HuBERT gives rise to syllabic organization without relying on external labels or modalities, and potentially provides novel data-driven units for spoken language modeling.

Self-Supervised Learning (SSL) based models of speech have shown remarkable performance on a range of downstream tasks. These state-of-the-art models have remained blackboxes, but many recent studies have begun “probing” models like HuBERT, to correlate their internal representations to different aspects of speech. In this paper, we show “inference of articulatory kinematics” as fundamental property of SSL models, i.e., the ability of these models to transform acoustics into the causal articulatory dynamics underlying the speech signal. We also show that this abstraction is largely overlapping across the language of the data used to train the model, with preference to the language with similar phonological system. Furthermore, we show that with simple affine transformations, Acoustic-to-Articulatory inversion (AAI) is transferrable across speakers, even across genders, languages, and dialects, showing the generalizability of this property. Together, these results shed new light on the internals of SSL models that are critical to their superior performance, and open up new avenues into language-agnostic universal models for speech engineering, that are interpretable and grounded in speech science.

Spoken Question Answering (SQA) is essential for machines to reply to user’s question by finding the answer span within a given spoken passage. SQA has been previously achieved without ASR to avoid recognition errors and Out-of-Vocabulary (OOV) problems. However, the real-world problem of Open-domain SQA (openSQA), in which the machine needs to first retrieve passages that possibly contain the answer from a spoken archive in addition, was never considered. This paper proposes the first known end-to-end frame-work, Speech Dense Passage Retriever (SpeechDPR), for the retrieval component of the openSQA problem. SpeechDPR learns a sentence-level semantic representation by distilling knowledge from the cascading model of unsupervised ASR (UASR) and text dense retriever (TDR). No manually transcribed speech data is needed. Initial experiments showed performance comparable to the cascading model of UASR and TDR, and significantly better when UASR was poor, verifying this approach is more robust to speech recognition errors.

Audio-visual representation learning aims to develop systems with human-like perception by utilizing correlation between auditory and visual information. However, current models often focus on a limited set of tasks, and generalization abilities of learned representations are unclear. To this end, we propose the AV-SUPERB benchmark that enables general-purpose evaluation of unimodal audio/visual and bimodal fusion representations on 7 datasets covering 5 audio-visual tasks in speech and audio processing. We evaluate 5 recent self-supervised models and show that none of these models generalize to all tasks, emphasizing the need for future study on improving universal model performance. In addition, we show that representations may be improved with intermediate-task fine-tuning and audio event classification with AudioSet serves as a strong intermediate task. We release our benchmark with evaluation code1and a model submission platform2to encourage further research in audio-visual learning.

The foundation model paradigm leverages a shared foundation model to achieve state-of-the-art (SOTA) performance for various tasks, requiring minimal downstream-specific data collection and modeling. This approach has proven crucial in the field of Natural Language Processing (NLP). However, the speech processing community lacks a similar setup to explore the paradigm systematically. To bridge this gap, we establish the Speech processing Universal PERformance Benchmark (SUPERB). SUPERB represents an ecosystem designed to evaluate foundation models across a wide range of speech processing tasks, facilitating the sharing of results on an online leaderboard and fostering collaboration through a community-driven benchmark database that aids in new development cycles. We present a unified learning framework for solving the speech processing tasks in SUPERB with the frozen foundation model followed by task-specialized lightweight prediction heads. Combining our results with community submissions, we verify that the framework is simple yet effective, as the best-performing foundation model shows competitive generalizability across most SUPERB tasks. Finally, we conduct a series of analyses to offer an in-depth understanding of SUPERB and speech foundation models, including information flows across tasks inside the models and the statistical significance and robustness of the benchmark.

The 2023 Multilingual Speech Universal Performance Benchmark (ML-SUPERB) Challenge expands upon the acclaimed SUPERB framework, emphasizing self-supervised models in multilingual speech recognition and language identification. The challenge comprises a research track focused on applying ML-SUPERB to specific multilingual subjects, a Challenge Track for model submissions, and a New Language Track where language resource researchers can contribute and evaluate their low-resource language data in the context of the latest progress in multilingual speech recognition. The challenge garnered 12 model submissions and 54 language corpora, resulting in a comprehensive benchmark encompassing 154 languages. The findings indicate that merely scaling models is not the definitive solution for multilingual speech tasks, and a variety of speech/voice types present significant challenges in multilingual speech processing.

Recent self-supervised learning (SSL) models have proven to learn rich representations of speech, which can readily be utilized by diverse downstream tasks. To understand such utilities, various analyses have been done for speech SSL models to reveal which and how information is encoded in the learned representations. Although the scope of previous analyses is extensive in acoustic, phonetic, and semantic perspectives, the physical grounding by speech production has not yet received full attention. To bridge this gap, we conduct a comprehensive analysis to link speech representations to articulatory trajectories measured by electromagnetic articulography (EMA). Our analysis is based on a linear probing approach where we measure articulatory score as an average correlation of linear mapping to EMA. We analyze a set of SSL models selected from the leaderboard of the SUPERB benchmark [1] and perform further layer-wise analyses on two most successful models, Wav2Vec 2.0 [2] and HuBERT [3]. Surprisingly, representations from the recent speech SSL models are highly correlated with EMA traces (best: r =0.81), and only 5 minutes are sufficient to train a linear model with high performance (r =0.77). Our findings suggest that SSL models learn to align closely with continuous articulations, and provide a novel insight into speech SSL.

Automatic speech recognition research focuses on training and evaluating on static datasets. Yet, as speech models are increasingly deployed on personal devices, such models encounter user-specific distributional shifts. To simulate this real-world scenario, we introduce LibriContinual, a continual learning benchmark for speaker-specific domain adaptation derived from LibriVox audiobooks, with data corresponding to 118 individual speakers and 6 train splits per speaker of different sizes. Additionally, current speech recognition models and continual learning algorithms are not optimized to be compute-efficient. We adapt a general-purpose training algorithm NetAug for ASR and create a novel Conformer variant called the DisConformer (Disentangled Conformer). This algorithm produces ASR models consisting of a frozen ‘core’ network for general-purpose use and several tunable ‘augment’ networks for speaker-specific tuning. Using such models, we propose a novel compute-efficient continual learning algorithm called DisentangledCL. Our experiments show that the DisConformer models significantly outperform base-lines on general ASR i.e. LibriSpeech (15.58% rel. WER on test-other). On speaker-specific LibriContinual they significantly outper-form trainable-parameter-matched baselines (by 20.65% rel. WER on test) and even match fully finetuned baselines in some settings.

The research community has produced many successful self-supervised speech representation learning methods over the past few years. Discrete units have been utilized in various self-supervised learning frameworks, such as VQ-VAE [1], wav2vec 2.0 [2], Hu-BERT [3], and Wav2Seq [4]. This paper studies the impact of altering the granularity and improving the quality of these discrete acoustic units for pre-training encoder-only and encoder-decoder models. We systematically study the current proposals of using Byte-Pair Encoding (BPE) and new extensions that use cluster smoothing and Brown clustering. The quality of learned units is studied intrinsically using zero speech metrics and on the down-stream speech recognition (ASR) task. Our results suggest that longer-range units are helpful for encoder-decoder pre-training; however, encoder-only masked-prediction models cannot yet benefit from self-supervised word-like targets.

End-to-end multilingual ASR has become more appealing because of several reasons such as simplifying the training and deployment process and positive performance transfer from high-resource to low-resource languages. However, scaling up the number of languages, total hours, and number of unique tokens is not a trivial task. This paper explores large-scale multilingual ASR models on 70 languages. We inspect two architectures: (1) Shared embedding and output and (2) Multiple embedding and output model. In the shared model experiments, we show the importance of tokenization strategy across different languages. Later, we use our optimal tokenization strategy to train multiple embedding and output model to further improve our result. Our multilingual ASR achieves 13.9%-15.6% average WER relative improvement compared to monolingual models. We show that our multilingual ASR generalizes well on an unseen dataset and domain, achieving 9.5% and 7.5% WER on Multilingual Librispeech (MLS) with zero-shot and finetuning, respectively.

In this paper, we show that representations capturing syllabic units emerge when training a self-supervised speech model with a visually-grounded training objective. We demonstrate that a nearly identical model architecture (HuBERT) trained with a masked language modeling loss does not exhibit this same ability, suggesting that the visual grounding objective is responsible for the emergence of this phenomenon. We propose the use of a minimum cut algorithm to automatically predict syllable boundaries in speech, followed by a 2-stage clustering method to group identical syllables together. We show that our model not only outperforms a state-of-the-art syllabic segmentation method on the language it was trained on (English), but also generalizes in a zero-shot fashion to Estonian. Finally, we show that the same model is capable of zero-shot generalization for a word segmentation task on 4 other languages from the Zerospeech Challenge, in some cases beating the previous state-of-the-art.

Abstract We introduce dGSLM, the first “textless” model able to generate audio samples of naturalistic spoken dialogues. It uses recent work on unsupervised spoken unit discovery coupled with a dual-tower transformer architecture with cross-attention trained on 2000 hours of two-channel raw conversational audio (Fisher dataset) without any text or labels. We show that our model is able to generate speech, laughter, and other paralinguistic signals in the two channels simultaneously and reproduces more naturalistic and fluid turn taking compared to a text-based cascaded model.1,2

State-of-the-art encoder-decoder models (e.g. for machine translation (MT) or automatic speech recognition (ASR)) are constructed and trained end-to-end as an atomic unit. No component of the model can be (re-)used without the others, making it impossible to share parts, e.g. a high resourced decoder, across tasks. We describe LegoNN, a procedure for building encoder-decoder architectures in a way so that its parts can be applied to other tasks without the need for any fine-tuning. To achieve this reusability, the interface between encoder and decoder modules is grounded to a sequence of marginal distributions over a pre-defined discrete vocabulary. We present two approaches for ingesting these marginals; one is differentiable, allowing the flow of gradients across the entire network, and the other is gradient-isolating. To enable the portability of decoder modules between MT tasks for different source languages and across other tasks like ASR, we introduce a modality agnostic encoder which consists of a length control mechanism to dynamically adapt encoders' output lengths in order to match the expected input length range of pre-trained decoders. We present several experiments to demonstrate the effectiveness of LegoNN models: a trained language generation LegoNN decoder module from German-English (De-En) MT task can be reused without any fine-tuning for the Europarl English ASR and the Romanian-English (Ro-En) MT tasks, matching or beating the performance of baseline. After fine-tuning, LegoNN models improve the Ro-En MT task by 1.5 BLEU points and achieve 12.5% relative WER reduction on the Europarl ASR task. To show how the approach generalizes, we compose a LegoNN ASR model from three modules – each has been learned within different end-to-end trained models on three different datasets – achieving an overall WER reduction of 19.5%.

Eugene Kharitonov, Ann Lee, Adam Polyak, Yossi Adi, Jade Copet, Kushal Lakhotia, Tu Anh Nguyen, Morgane Riviere, Abdelrahman Mohamed, Emmanuel Dupoux, Wei-Ning Hsu. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Yun Tang, Hongyu Gong, Ning Dong, Changhan Wang, Wei-Ning Hsu, Jiatao Gu, Alexei Baevski, Xian Li, Abdelrahman Mohamed, Michael Auli, Juan Pino. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Hsiang-Sheng Tsai, Heng-Jui Chang, Wen-Chin Huang, Zili Huang, Kushal Lakhotia, Shu-wen Yang, Shuyan Dong, Andy Liu, Cheng-I Lai, Jiatong Shi, Xuankai Chang, Phil Hall, Hsuan-Jui Chen, Shang-Wen Li, Shinji Watanabe, Abdelrahman Mohamed, Hung-yi Lee. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Felix Kreuk, Adam Polyak, Jade Copet, Eugene Kharitonov, Tu Anh Nguyen, Morgan Rivière, Wei-Ning Hsu, Abdelrahman Mohamed, Emmanuel Dupoux, Yossi Adi. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

We consider two federated learning algorithms for training partially personalized models, where the shared and personal parameters are updated either simultaneously or alternately on the devices. Both algorithms have been proposed in the literature, but their convergence properties are not fully understood, especially for the alternating variant. We provide convergence analyses of both algorithms in the general nonconvex setting with partial participation and delineate the regime where one dominates the other. Our experiments on real-world image, text, and speech datasets demonstrate that (a) partial personalization can obtain most of the benefits of full model personalization with a small fraction of personal parameters, and, (b) the alternating update algorithm outperforms the simultaneous update algorithm by a small but consistent margin.

Spoken Question Answering (SQA) is to find the answer from a spoken document given a question, which is crucial for personal assistants when replying to the queries from the users.Existing SQA methods all rely on Automatic Speech Recognition (ASR) transcripts.Not only does ASR need to be trained with massive annotated data that are time and cost-prohibitive to collect for low-resourced languages, but more importantly, very often the answers to the questions include name entities or out-of-vocabulary words that cannot be recognized correctly.Also, ASR aims to minimize recognition errors equally over all words, including many function words irrelevant to the SQA task.Therefore, SQA without ASR transcripts (textless) is always highly desired, although known to be very difficult.This work proposes Discrete Spoken Unit Adaptive Learning (DUAL), leveraging unlabeled data for pre-training and finetuned by the SQA downstream task.The time intervals of spoken answers can be directly predicted from spoken documents.We also release a new SQA benchmark corpus, NMSQA, for data with more realistic scenarios.We empirically showed that DUAL yields results comparable to those obtained by cascading ASR and text QA model and robust to real-world data.Our code and model will be open-sourced 1 .

Audio-based automatic speech recognition (ASR) degrades significantly in noisy environments and is particularly vulnerable to interfering speech, as the model cannot determine which speaker to transcribe.Audio-visual speech recognition (AVSR) systems improve robustness by complementing the audio stream with the visual information that is invariant to noise and helps the model focus on the desired speaker.However, previous AVSR work focused solely on the supervised learning setup; hence the progress was hindered by the amount of labeled data available.In this work, we present a self-supervised AVSR framework built upon Audio-Visual HuBERT (AV-HuBERT), a state-of-theart audio-visual speech representation learning model.On the largest available AVSR benchmark dataset LRS3, our approach outperforms prior state-of-the-art by ∼ 50% (28.0% vs. 14.1%) using less than 10% of labeled data (433hr vs. 30hr) in the presence of babble noise, while reducing the WER of an audio-based model by over 75% (25.8% vs. 5.8%) on average 1 .

This paper investigates self-supervised pre-training for audiovisual speaker representation learning where a visual stream showing the speaker's mouth area is used alongside speech as inputs.Our study focuses on the Audio-Visual Hidden Unit BERT (AV-HuBERT) approach, a recently developed generalpurpose audio-visual speech pre-training framework.We conducted extensive experiments probing the effectiveness of pretraining and visual modality.Experimental results suggest that AV-HuBERT generalizes decently to speaker related downstream tasks, improving label efficiency by roughly ten fold for both audio-only and audio-visual speaker verification.We also show that incorporating visual information, even just the lip area, greatly improves the performance and noise robustness, reducing EER by 38% in the clean condition and 75% in noisy conditions 1 .

With 4.5 million hours of English speech from 10 different sources across 120 countries and models of up to 10 billion parameters, we explore the frontiers of scale for automatic speech recognition.We propose data selection techniques to efficiently scale training data to find the most valuable samples in massive datasets.To efficiently scale model sizes, we leverage various optimizations such as sparse transducer loss and model sharding.By training 1-10B parameter universal English ASR models, we push the limits of speech recognition performance across many domains.Furthermore, our models learn powerful speech representations with zero and few-shot capabilities on novel domains and styles of speech, exceeding previous results across multiple in-house and public benchmarks.For speakers with disorders due to brain damage, our best zero-shot and few-shot models achieve 22% and 60% relative improvement on the AphasiaBank test set, respectively, while realizing the best performance on public social media videos.Furthermore, the same universal model reaches equivalent performance with 500x less in-domain data on the SPGISpeech financial-domain dataset.

We present the SUPERB challenge at SLT 2022, which aims at learning self-supervised speech representation for better performance, generalization, and efficiency. The challenge builds upon the SUPERB benchmark and implements metrics to measure the computation requirements of self-supervised learning (SSL) representation and to evaluate its generalizability and performance across the diverse SUPERB tasks. The SUPERB benchmark provides comprehensive coverage of popular speech processing tasks, from speech and speaker recognition to audio generation and semantic understanding. As SSL has gained interest in the speech community and showed promising outcomes, we envision the challenge to uplevel the impact of SSL techniques by motivating more practical designs of techniques beyond task performance. We summarize the results of 14 submitted models in this paper. We also discuss the main findings from those submissions and the future directions of SSL research.

End-to-end spoken language understanding (SLU) predicts intent directly from audio using a single model. It promises to improve the performance of assistant systems by leveraging acoustic information lost in the intermediate textual representation and preventing cascading errors from Automatic Speech Recognition (ASR). Further, having one unified model has efficiency advantages when deploying assistant systems on-device. However, the limited number of public audio datasets with semantic parse labels hinders the research progress in this area. In this paper, we release the Spoken Task-Oriented semantic Parsing (STOP) dataset1, the largest and most complex SLU dataset publicly available. Additionally, we define low-resource splits to establish a benchmark for improving SLU when limited labeled data is available. Furthermore, in addition to the human-recorded audio, we are releasing a TTS-generated versions to benchmark the performance for low-resource and domain adaptation of end-to-end SLU systems.

Abstract Finding word boundaries in continuous speech is challenging as there is little or no equivalent of a ‘space’ delimiter between words. Popular Bayesian non-parametric models for text segmentation (Goldwater et al., 2006, 2009) use a Dirichlet process to jointly segment sentences and build a lexicon of word types. We introduce DP-Parse, which uses similar principles but only relies on an instance lexicon of word tokens, avoiding the clustering errors that arise with a lexicon of word types. On the Zero Resource Speech Benchmark 2017, our model sets a new speech segmentation state-of-the-art in 5 languages. The algorithm monotonically improves with better input representations, achieving yet higher scores when fed with weakly supervised inputs. Despite lacking a type lexicon, DP-Parse can be pipelined to a language model and learn semantic and syntactic representations as assessed by a new spoken word embedding benchmark. 1

In this paper, we introduce the Kaizen framework that uses a continuously improving teacher to generate pseudo-labels for semi-supervised speech recognition (ASR). The proposed approach uses a teacher model which is updated as the exponential moving average (EMA) of the student model parameters. We demonstrate that it is critical for EMA to be accumulated with full-precision floating point. The Kaizen framework can be seen as a continuous version of the iterative pseudo-labeling approach for semi-supervised training. It is applicable for different training criteria, and in this paper we demonstrate its effectiveness for frame-level hybrid hidden Markov model-deep neural network (HMM-DNN) systems as well as sequence-level Connectionist Temporal Classification (CTC) based models. For large scale real-world unsupervised public videos in UK English and Italian languages the proposed approach i) shows more than 10% relative word error rate (WER) reduction over standard teacher-student training; ii) using just 10 hours of supervised data and a large amount of unsupervised data closes the gap to the upper-bound supervised ASR system that uses 650h or 2700h respectively.

Compared to vision and language applications, self-supervised pre-training approaches for ASR are challenged by three unique problems: (1) There are multiple sound units in each input utterance, (2) With audio-only pre-training, there is no lexicon of sound units, and (3) Sound units have variable lengths with no explicit segmentation. In this paper, we propose the Hidden-Unit BERT (HUBERT) model which utilizes a cheap k-means clustering step to provide aligned target labels for pre-training of a BERT model. A key ingredient of our approach is applying the predictive loss over the masked regions only. This allows the pre-training stage to benefit from the consistency of the unsupervised teacher rather that its intrinsic quality. Starting with a simple k-means teacher of 100 cluster, and using two iterations of clustering, the HUBERT model matches the state-of-the-art wav2vec 2.0 performance on the ultra low-resource Libri-light 10h, 1h, 10min supervised subsets.

Pseudo-labeling is the most adopted method for pre-training automatic speech recognition (ASR) models. However, its performance suffers with degrading quality of the supervised teacher model.Inspired by the successes of contrastive representation learning for both computer vision and speech applications, and more recently for supervised learning of visual objects [1], we propose Contrastive Semi-supervised Learning (CSL). CSL eschews directly predicting teacher generated pseudo-labels in favor of utilizing them to select positive and negative examples.In the challenging task of transcribing public social media videos, using CSL reduces the WER by 8%, compared to the standard Cross-Entropy pseudo-labeling (CE-PL), when 10hr of supervised data is used to annotate 75,000hr of videos. The WER reduction jumps to 19% under the ultra low-resource condition of using 1hr labels for teacher supervision. In out-of-domain conditions, CSL generalizes much better showing up to 17% WER reduction compared to the strongest CE-PL pre-trained model.

This paper presents XLSR which learns cross-lingual speech representations by pretraining a single model from the raw waveform of speech in multiple languages. We build on wav2vec 2.0 which is trained by solving a contrastive task over masked latent speech representations and jointly learns a quantization of the latents shared across languages. The resulting model is fine-tuned on labeled data and experiments show that cross-lingual pretraining significantly outperforms monolingual pretraining. On the CommonVoice benchmark, XLSR shows a relative phoneme error rate reduction of 72% compared to the best known results. On BABEL, our approach improves word error rate by 16% relative compared to a comparable system. Our approach enables a single multilingual speech recognition model which is competitive to strong individual models. Analysis shows that the latent discrete speech representations are shared across languages with increased sharing for related languages. We hope to catalyze research in low-resource speech understanding by releasing XLSR-53, a large model pretrained in 53 languages.

We propose using self-supervised discrete representations for the task of speech resynthesis. To generate disentangled representation, we separately extract low-bitrate representations for speech content, prosodic information, and speaker identity. This allows to synthesize speech in a controllable manner. We analyze various state-of-the-art, self-supervised representation learning methods and shed light on the advantages of each method while considering reconstruction quality and disentanglement properties. Specifically, we evaluate the F0 reconstruction, speaker identification performance (for both resynthesis and voice conversion), recordings' intelligibility, and overall quality using subjective human evaluation. Lastly, we demonstrate how these representations can be used for an ultra-lightweight speech codec. Using the obtained representations, we can get to a rate of 365 bits per second while providing better speech quality than the baseline methods. Audio samples can be found under the following link: speechbot.github.io/resynthesis.

Self-supervised learning (SSL) has proven vital for advancing research in natural language processing (NLP) and computer vision (CV).The paradigm pretrains a shared model on large volumes of unlabeled data and achieves state-of-the-art (SOTA) for various tasks with minimal adaptation.However, the speech processing community lacks a similar setup to systematically explore the paradigm.To bridge this gap, we introduce Speech processing Universal PERformance Benchmark (SUPERB).SUPERB is a leaderboard to benchmark the performance of a shared model across a wide range of speech processing tasks with minimal architecture changes and labeled data.Among multiple usages of the shared model, we especially focus on extracting the representation learned from SSL for its preferable re-usability.We present a simple framework to solve SUPERB tasks by learning task-specialized lightweight prediction heads on top of the frozen shared model.Our results demonstrate that the framework is promising as SSL representations show competitive generalizability and accessibility across SUPERB tasks.We release SUPERB as a challenge with a leaderboard 1 and a benchmark toolkit 2 to fuel the research in representation learning and general speech processing.

Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we propose the Hidden-Unit BERT (HuBERT) approach for self-supervised speech representation learning, which utilizes an offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our approach is applying the prediction loss over the masked regions only, which forces the model to learn a combined acoustic and language model over the continuous inputs. HuBERT relies primarily on the consistency of the unsupervised clustering step rather than the intrinsic quality of the assigned cluster labels. Starting with a simple k-means teacher of 100 clusters, and using two iterations of clustering, the HuBERT model either matches or improves upon the state-of-the-art wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h, 10h, 100h, and 960h fine-tuning subsets. Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets.

Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, Luke Zettlemoyer. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.

We compare self-supervised representation learning algorithms which either explicitly quantize the audio data or learn representations without quantization. We find the former to be more accurate since it builds a good vocabulary of the data through vq-wav2vec [1] self-supervision approach to enable learning of effective representations in subsequent BERT training. Different to previous work, we directly fine-tune the pre-trained BERT models on transcribed speech using a Connectionist Temporal Classification (CTC) loss instead of feeding the representations into a task-specific model. We also propose a BERT-style model learning directly from the continuous audio data and compare pre-training on raw audio to spectral features. Fine-tuning a BERT model on 10 hour of labeled Librispeech data with a vq-wav2vec vocabulary is almost as good as the best known reported system trained on 100 hours of labeled data on test-clean, while achieving a 25% WER reduction on test-other. When using only 10 minutes of labeled data, WER is 25.2 on test-other and 16.3 on test-clean. This demonstrates that self-supervision can enable speech recognition systems trained on a near-zero amount of transcribed data.

We introduce a new collection of spoken English audio suitable for training speech recognition systems under limited or no supervision. It is derived from open-source audio books from the LibriVox project. It contains over 60K hours of audio, which is, to our knowledge, the largest freely-available corpus of speech. The audio has been segmented using voice activity detection and is tagged with SNR, speaker ID and genre descriptions. Additionally, we provide baseline systems and evaluation metrics working under three settings: (1) the zero resource/unsupervised setting (ABX), (2) the semi- supervised setting (PER, CER) and (3) the distant supervision setting (WER). Settings (2) and (3) use limited textual resources (10 minutes to 10 hours) aligned with the speech. Setting (3) uses large amounts of unaligned text. They are evaluated on the standard LibriSpeech dev and test sets for comparison with the supervised state-of-the-art.

Supervised ASR models have reached unprecedented levels of accuracy, thanks in part to ever-increasing amounts of labelled training data. However, in many applications and locales, only moderate amounts of data are available, which has led to a surge in semi- and weakly-supervised learning research. In this paper, we conduct a large-scale study evaluating the effectiveness of weakly-supervised learning for speech recognition by using loosely related contextual information as a surrogate for ground-truth labels. For weakly supervised training, we use 50k hours of public English social media videos along with their respective titles and post text to train an encoder-decoder transformer model. Our best encoder-decoder models achieve an average of 20.8% WER reduction over a 1000 hours supervised baseline, and an average of 13.4% WER reduction when using only the weakly supervised encoder for CTC fine-tuning. Our results show that our setup for weak supervision improved both the encoder acoustic representations as well as the decoder language generation abilities.

We propose and evaluate transformer-based acoustic models (AMs) for hybrid speech recognition. Several modeling choices are discussed in this work, including various positional embedding methods and an iterated loss to enable training deep transformers. We also present a preliminary study of using limited right context in transformer models, which makes it possible for streaming applications. We demonstrate that on the widely used Librispeech benchmark, our transformer-based AM outperforms the best published hybrid result by 19% to 26% relative when the standard n-gram language model (LM) is used. Combined with neural network LM for rescoring, our proposed approach achieves state-of-the-art results on Librispeech. Our findings are also confirmed on a much larger internal dataset.

Many semi- and weakly-supervised approaches have been investigated for overcoming the labeling cost of building high quality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two self-labeling methods on one hand, and weakly-supervised pretraining using contextual metadata on the other. We investigate distillation methods at the frame level and the sequence level for hybrid, encoder-only CTC-based, and encoder-decoder speech recognition systems on Dutch and Romanian languages using 27,000 and 58,000 hours of unlabeled audio respectively. Although all approaches improved upon their respective baseline WERs by more than 8%, sequence-level distillation for encoder-decoder models provided the largest relative WER reduction of 20% compared to the strongest data-augmented supervised baseline.

We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.

Recent advances in spoken language technologies and the introduction of many customer facing products, have given rise to a wide customer reliance on smart personal assistants for many of their daily tasks. In this paper, we present a system to reduce users' cognitive load by extending personal assistants with long-term personal memory where users can store and retrieve by voice, arbitrary pieces of information. The problem is framed as a neural retrieval based question answering system where answers are selected from previously stored user memories. We propose to directly optimize the end-to-end retrieval performance, measured by the F1-score, using reinforcement learning, leading to better performance on our experimental test set(s).

The problem of automatically generating a computer program from some specification has been studied since the early days of AI. Recently, two competing approaches for `automatic program learning’ have received significant attention: (1) `neural program synthesis’, where a neural network is conditioned on input/output (I/O) examples and learns to generate a program, and (2) `neural program induction’, where a neural network generates new outputs directly using a latent program representation. Here, for the first time, we directly compare both approaches on a large-scale, real-world learning task and we additionally contrast to rule-based program synthesis, which uses hand-crafted semantics to guide the program generation. Our neural models use a modified attention RNN to allow encoding of variable-sized sets of I/O pairs, which achieve 92\% accuracy on a real-world test set, compared to the 34\% accuracy of the previous best neural synthesis approach. The synthesis model also outperforms a comparable induction model on this task, but we more importantly demonstrate that the strength of each approach is highly dependent on the evaluation metric and end-user application. Finally, we show that we can train our neural models to remain very robust to the type of noise expected in real-world data (e.g., typos), while a highly-engineered rule-based system fails entirely.

Segmental structure is a common pattern in many types of sequences such as phrases in human languages. In this paper, we present a probabilistic model for sequences via their segmentations. The probability of a segmented sequence is calculated as the product of the probabilities of all its segments, where each segment is modeled using existing tools such as recurrent neural networks. Since the segmentation of a sequence is usually unknown in advance, we sum over all valid segmentations to obtain the final probability for the sequence. An efficient dynamic programming algorithm is developed for forward and backward computations without resorting to any approximation. We demonstrate our approach on text segmentation and speech recognition tasks. In addition to quantitative results, we also show that our approach can discover meaningful segments in their respective application contexts.

Long short-term memory (LSTM) recurrent neural networks (RNNs) have recently shown significant performance improvements over deep feed-forward neural networks. A key aspect of these models is the use of time recurrence, combined with a gating architecture that allows them to track the long-term dynamics of speech. Inspired by human spectrogram reading, we recently proposed the frequency LSTM (F-LSTM) that performs 1-D recurrence over the frequency axis and then performs 1-D recurrence over the time axis. In this study, we further improve the acoustic model by proposing a 2-D, time-frequency (TF) LSTM. The TF-LSTM jointly scans the input over the time and frequency axes to model spectro-temporal warping, and then uses the output activations as the input to a time LSTM (T-LSTM). The joint time-frequency modeling better normalizes the features for the upper layer T-LSTMs. Evaluated on a 375-hour short message dictation task, the proposed TF-LSTM obtained a 3.4% relative WER reduction over the best T-LSTM. The invariance property achieved by joint time-frequency analysis is demonstrated on a mismatched test set, where the TF-LSTM achieves a 14.2% relative WER reduction over the best T-LSTM.

Deep neural networks have achieved great successes on various machine learning tasks, however, there are many open fundamental questions to be answered. In this paper, we tackle the problem of quantifying the quality of learned wights of different networks with possibly different architectures, going beyond considering the final classification error as the only metric. We introduce \emphExtended Data Jacobian Matrix to help analyze properties of networks of various structures, finding that, the spectrum of the extended data jacobian matrix is a strong discriminating factor for networks of different structures and performance. Based on such observation, we propose a novel regularization method, which manages to improve the network performance comparably to dropout, which in turn verifies the observation.

Long short-term memory (LSTM) recurrent neural networks (RNNs) have recently shown significant performance improvements over deep feed-forward neural networks (DNNs). A key aspect of these models is the use of time recurrence, combined with a gating architecture that ameliorates the vanishing gradient problem. Inspired by human spectrogram reading, in this paper we propose an extension to LSTMs that performs the recurrence in frequency as well as in time. This model first scans the frequency bands to generate a summary of the spectral information, and then uses the output layer activations as the input to a traditional time LSTM (T-LSTM). Evaluated on a Microsoft short message dictation task, the proposed model obtained a 3.6% relative word error rate reduction over the T-LSTM.

Long short-term memory (LSTM) acoustic models have recently achieved state-of-the-art results on speech recognition tasks. As a type of recurrent neural network, LSTMs potentially have the ability to model long-span phenomena relating the spectral input to linguistic units. However, it has not been clear whether their observed performance is actually due to this capability, or instead if it is due to a better modeling of short term dynamics through the recurrence. In this paper. we answer this question by applying a windowed (truncated) LSTM to conversational speech transcription, and find that a limited context is adequate, and that it is not necessaary to scan the entire utterance. The sliding window approach allows not only incremental (online) recognition with a bidirectional model, but also frame-wise randomization (as opposed to utterance randomization), which results in faster convergence. On the SWBD/Fisher corpus, applying bidirectional LSTM RNNs to spectral windows of about 0.5s improves WER on the Hub5'00 benchmark set by 16% relative compared to our best sequence-trained DNN. On an extended 3850h training set that that also includes lectures, the relative gain becomes 28% (Hub5'00 WER 9.2%). In-house conversational data improves by 12 to 17% relative.

Many features used in speech recognition tasks are hand-crafted and are not always related to the objective at hand, that is minimizing word error rate. Recently, we showed that replacing a perceptually motivated mel-filter bank with a filter bank layer that is learned jointly with the rest of a deep neural network was promising. In this paper, we extend filter learning to a speaker-adapted, state-of-the-art system. First, we incorporate delta learning into the filter learning framework. Second, we incorporate various speaker adaptation techniques, including VTLN warping and speaker identity features. On a 50-hour English Broadcast News task, we show that we can achieve a 5% relative improvement in word error rate (WER) using the filter and delta learning, compared to having a fixed set of filters and deltas. Furthermore, after speaker adaptation, we find that filter and delta learning allows for a 3% relative improvement in WER compared to a state-of-the-art CNN.

Recently, the hybrid deep neural network (DNN)-hidden Markov model (HMM) has been shown to significantly improve speech recognition performance over the conventional Gaussian mixture model (GMM)-HMM. The performance improvement is partially attributed to the ability of the DNN to model complex correlations in speech features. In this paper, we show that further error rate reduction can be obtained by using convolutional neural networks (CNNs). We first present a concise description of the basic CNN and explain how it can be used for speech recognition. We further propose a limited-weight-sharing scheme that can better model speech features. The special structure such as local connectivity, weight sharing, and pooling in CNNs exhibits some degree of invariance to small shifts of speech features along the frequency axis, which is important to deal with speaker and environment variations. Experimental results show that CNNs reduce the error rate by 6%-10% compared with DNNs on the TIMIT phone recognition and the voice search large vocabulary speech recognition tasks.

Deep Bidirectional LSTM (DBLSTM) recurrent neural networks have recently been shown to give state-of-the-art performance on the TIMIT speech database. However, the results in that work relied on recurrent-neural-network-specific objective functions, which are difficult to integrate with existing large vocabulary speech recognition systems. This paper investigates the use of DBLSTM as an acoustic model in a standard neural network-HMM hybrid system. We find that a DBLSTM-HMM hybrid gives equally good results on TIMIT as the previous work. It also outperforms both GMM and deep network benchmarks on a subset of the Wall Street Journal corpus. However the improvement in word error rate over the deep network is modest, despite a great increase in framelevel accuracy. We conclude that the hybrid approach with DBLSTM appears to be well suited for tasks where acoustic modelling predominates. Further investigation needs to be conducted to understand how to better leverage the improvements in frame-level accuracy towards better word error rates.

Deep Convolutional Neural Networks (CNNs) are more powerful than Deep Neural Networks (DNN), as they are able to better reduce spectral variation in the input signal. This has also been confirmed experimentally, with CNNs showing improvements in word error rate (WER) between 4-12% relative compared to DNNs across a variety of LVCSR tasks. In this paper, we describe different methods to further improve CNN performance. First, we conduct a deep analysis comparing limited weight sharing and full weight sharing with state-of-the-art features. Second, we apply various pooling strategies that have shown improvements in computer vision to an LVCSR speech task. Third, we introduce a method to effectively incorporate speaker adaptation, namely fMLLR, into log-mel features. Fourth, we introduce an effective strategy to use dropout during Hessian-free sequence training. We find that with these improvements, particularly with fMLLR and dropout, we are able to achieve an additional 2-3% relative improvement in WER on a 50-hour Broadcast News task over our previous best CNN baseline. On a larger 400-hour BN task, we find an additional 4-5% relative improvement over our previous best CNN baseline.

Mel-filter banks are commonly used in speech recognition, as they are motivated from theory related to speech production and perception. While features derived from mel-filter banks are quite popular, we argue that this filter bank is not really an appropriate choice as it is not learned for the objective at hand, i.e. speech recognition. In this paper, we explore replacing the filter bank with a filter bank layer that is learned jointly with the rest of a deep neural network. Thus, the filter bank is learned to minimize cross-entropy, which is more closely tied to the speech recognition objective. On a 50-hour English Broadcast News task, we show that we can achieve a 5% relative improvement in word error rate (WER) using the filter bank learning approach, compared to having a fixed set of filters.

Recurrent neural networks (RNNs) are a powerful model for sequential data. End-to-end training methods such as Connectionist Temporal Classification make it possible to train RNNs for sequence labelling problems where the input-output alignment is unknown. The combination of these methods with the Long Short-term Memory RNN architecture has proved particularly fruitful, delivering state-of-the-art results in cursive handwriting recognition. However RNN performance in speech recognition has so far been disappointing, with better results returned by deep feedforward networks. This paper investigates deep recurrent neural networks, which combine the multiple levels of representation that have proved so effective in deep networks with the flexible use of long range context that empowers RNNs. When trained end-to-end with suitable regularisation, we find that deep Long Short-term Memory RNNs achieve a test set error of 17.7% on the TIMIT phoneme recognition benchmark, which to our knowledge is the best recorded score.

Convolutional Neural Networks (CNNs) are an alternative type of neural network that can be used to reduce spectral variations and model spectral correlations which exist in signals. Since speech signals exhibit both of these properties, CNNs are a more effective model for speech compared to Deep Neural Networks (DNNs). In this paper, we explore applying CNNs to large vocabulary speech tasks. First, we determine the appropriate architecture to make CNNs effective compared to DNNs for LVCSR tasks. Specifically, we focus on how many convolutional layers are needed, what is the optimal number of hidden units, what is the best pooling strategy, and the best input feature type for CNNs. We then explore the behavior of neural network features extracted from CNNs on a variety of LVCSR tasks, comparing CNNs to DNNs and GMMs. We find that CNNs offer between a 13-30% relative improvement over GMMs, and a 4-12% relative improvement over DNNs, on a 400-hr Broadcast News and 300-hr Switchboard task.

Convolutional Neural Networks (CNN) have showed success in achieving translation invariance for many image processing tasks. The success is largely attributed to the use of local filtering and max-pooling in the CNN architecture. In this paper, we propose to apply CNN to speech recognition within the framework of hybrid NN-HMM model. We propose to use local filtering and max-pooling in frequency domain to normalize speaker variance to achieve higher multi-speaker speech recognition performance. In our method, a pair of local filtering layer and max-pooling layer is added at the lowest end of neural network (NN) to normalize spectral variations of speech signals. In our experiments, the proposed CNN architecture is evaluated in a speaker independent speech recognition task using the standard TIMIT data sets. Experimental results show that the proposed CNN method can achieve over 10% relative error reduction in the core TIMIT test sets when comparing with a regular NN using the same number of hidden layers and weights. Our results also show that the best result of the proposed CNN model is better than previously published results on the same TIMIT test sets that use a pre-trained deep NN model.

Deep Belief Networks (DBNs) are a very competitive alternative to Gaussian mixture models for relating states of a hidden Markov model to frames of coefficients derived from the acoustic input. They are competitive for three reasons: DBNs can be fine-tuned as neural networks; DBNs have many non-linear hidden layers; and DBNs are generatively pre-trained. This paper illustrates how each of these three aspects contributes to the DBN's good recognition performance using both phone recognition performance on the TIMIT corpus and a dimensionally reduced visualization of the relationships between the feature vectors learned by the DBNs that preserves the similarity structure of the feature vectors at multiple scales. The same two methods are also used to investigate the most suitable type of input representation for a DBN.

Gaussian mixture models are currently the dominant technique for modeling the emission distribution of hidden Markov models for speech recognition. We show that better phone recognition on the TIMIT dataset can be achieved by replacing Gaussian mixture models by deep neural networks that contain many layers of features and a very large number of parameters. These networks are first pre-trained as a multi-layer generative model of a window of spectral feature vectors without making use of any discriminative information. Once the generative pre-training has designed the features, we perform discriminative fine-tuning using backpropagation to adjust the features slightly to make them better at predicting a probability distribution over the states of monophone hidden Markov models.

To date, there has been limited work in applying Deep Belief Networks (DBNs) for acoustic modeling in LVCSR tasks, with past work using standard speech features. However, a typical LVCSR system makes use of both feature and model-space speaker adaptation and discriminative training. This paper explores the performance of DBNs in a state-of-the-art LVCSR system, showing improvements over Multi-Layer Perceptrons (MLPs) and GMM/HMMs across a variety of features on an English Broadcast News task. In addition, we provide a recipe for data parallelization of DBN training, showing that data parallelization can provide linear speed-up in the number of machines, without impacting WER.

Deep Belief Networks (DBNs) are multi-layer generative models. They can be trained to model windows of coefficients extracted from speech and they discover multiple layers of features that capture the higher-order statistical structure of the data. These features can be used to initialize the hidden units of a feed-forward neural network that is then trained to predict the HMM state for the central frame of the window. Initializing with features that are good at generating speech makes the neural network perform much better than initializing with random weights. DBNs have already been used successfully for phone recognition with input coefficients that are MFCCs or filterbank outputs. In this paper, we demonstrate that they work even better when their inputs are speaker adaptive, discriminative features. On the standard TIMIT corpus, they give phone error rates of 19.6% using monophone HMMs and a bigram language model and 19.4% using monophone HMMs and a trigram language model.

For decades, Hidden Markov Models (HMMs) have been the state-of-the-art technique for acoustic modeling despite their unrealistic independence assumptions and the very limited representational capacity of their hidden states. Conditional Restricted Boltzmann Machines (CRBMs) have recently proved to be very effective for modeling motion capture sequences and this paper investigates the application of this more powerful type of generative model to acoustic modeling. On the standard TIMIT corpus, one type of CRBM outperforms HMMs and is comparable with the best other methods, achieving a phone error rate (PER) of 26.7% on the TIMIT core test set.

This paper reports our recent exploration of the layer-by-layer learning strategy for training a multi-layer generative model of patches of speech spectrograms. The top layer of the generative model learns binary codes that can be used for efficient compression of speech and could also be used for scalable speech recognition or rapid speech content retrieval. Each layer of the generative model is fully connected to the layer below and the weights on these connections are pretrained efficiently by using the contrastive divergence approximation to the log likelihood gradient. After layer-bylayer pre-training we “unroll” the generative model to form a deep auto-encoder, whose parameters are then fine-tuned using back-propagation. To reconstruct the full-length speech spectrogram, individual spectrogram segments predicted by their respective binary codes are combined using an overlapand-add method. Experimental results on speech spectrogram coding demonstrate that the binary codes produce a logspectral distortion that is approximately 2 dB lower than a subband vector quantization technique over the entire frequency range of wide-band speech. Index Terms: deep learning, speech feature extraction, neural networks, auto-encoder, binary codes, Boltzmann machine

Recently, Deep Belief Networks (DBNs) have been proposed for phone recognition and were found to achieve highly competitive performance. In the original DBNs, only framelevel information was used for training DBN weights while it has been known for long that sequential or full-sequence information can be helpful in improving speech recognition accuracy. In this paper we investigate approaches to optimizing the DBN weights, state-to-state transition parameters, and language model scores using the sequential discriminative training criterion. We describe and analyze the proposed training algorithm and strategy, and discuss practical issues and how they affect the final results. We show that the DBNs learned using the sequence-based training criterion outperform those with frame-based criterion using both threelayer and six-layer models, but the optimization procedure for the deeper DBN is more difficult for the former criterion.

Straightforward application of Deep Belief Nets (DBNs) to acoustic modeling produces a rich distributed representation of speech data that is useful for recognition and yields impressive results on the speaker-independent TIMIT phone recognition task. However, the first-layer Gaussian-Bernoulli Restricted Boltzmann Machine (GRBM) has an important limitation, shared with mixtures of diagonal-covariance Gaussians: GRBMs treat different components of the acoustic input vector as conditionally independent given the hidden state. The mean-covariance restricted Boltzmann machine (mcRBM), first introduced for modeling natural images, is a much more representationally efficient and powerful way of modeling the covariance structure of speech data. Every configuration of the precision units of the mcRBM specifies a different precision matrix for the conditional distribution over the acoustic space. In this work, we use the mcRBM to learn features of speech data that serve as input into a standard DBN. The mcRBM features combined with DBNs allow us to achieve a phone error rate of 20.5\%, which is superior to all published results on speaker-independent TIMIT to date.