Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale

Large-scale generative models such as GPT and DALL-E have revolutionized natural language processing and computer vision research. These models not only generate high fidelity text or image outputs, but are also generalists which can solve tasks not explicitly taught. In contrast, speech generative models are still primitive in terms of scale and task generalization. In this paper, we present Voicebox, the most versatile text-guided generative model for speech at scale. Voicebox is a non-autoregressive flow-matching model trained to infill speech, given audio context and text, trained on over 50K hours of speech that are neither filtered nor enhanced. Similar to GPT, Voicebox can perform many different tasks through in-context learning, but is more flexible as it can also condition on future context. Voicebox can be used for mono or cross-lingual zero-shot text-to-speech synthesis, noise removal, content editing, style conversion, and diverse sample generation. In particular, Voicebox outperforms the state-of-the-art zero-shot TTS model VALL-E on both intelligibility (5.9% vs 1.9% word error rates) and audio similarity (0.580 vs 0.681) while being up to 20 times faster. See voicebox.metademolab.com for a demo of the model.

Self-Supervised Representations for Singing Voice Conversion

A singing voice conversion model converts a song in the voice of an arbitrary source singer to the voice of a target singer. Recently, methods that leverage self-supervised audio representations such as HuBERT and Wav2Vec 2.0 have helped further the state-of-the-art. Though these methods produce more natural and melodic singing outputs, they often rely on confusion and disentanglement losses to render the self-supervised representations speaker and pitch-invariant. In this paper, we circumvent disentanglement training and propose a new model that leverages ASR fine-tuned self-supervised representations as inputs to a HiFi-GAN neural vocoder for singing voice conversion. We experiment with different f0 encoding schemes and show that an f0 harmonic generation module that uses a parallel bank of transposed convolutions (PBTC) alongside ASR fine-tuned Wav2Vec 2.0 features results in the best singing voice conversion quality. Additionally, the model is capable of making a spoken voice sing. We also show that a simple f0 shifting scheme during inference helps retain singer identity and bolsters the performance of our singing voice conversion model. Our results are backed up by extensive MOS studies that compare different ablations and baselines.

Voice-preserving Zero-shot Multiple Accent Conversion

Most people who have tried to learn a foreign language would have experienced difficulties understanding or speaking with a native speaker's accent. For native speakers, understanding or speaking a new accent is likewise a difficult task. An accent conversion system that changes a speaker's accent but preserves that speaker's voice identity, such as timbre and pitch, has the potential for a range of applications, such as communication, language learning, and entertainment. Existing accent conversion models tend to change the speaker identity and accent at the same time. Here, we use adversarial learning to disentangle accent dependent features while retaining other acoustic characteristics. What sets our work apart from existing accent conversion models is the capability to convert an unseen speaker's utterance to multiple accents while preserving its original voice identity. Subjective evaluations show that our model generates audio that sound closer to the target accent and like the original speaker.

Towards zero-shot Text-based voice editing using acoustic context conditioning, utterance embeddings, and reference encoders

Text-based voice editing (TBVE) uses synthetic output from text-to-speech (TTS) systems to replace words in an original recording. Recent work has used neural models to produce edited speech that is similar to the original speech in terms of clarity, speaker identity, and prosody. However, one limitation of prior work is the usage of finetuning to optimise performance: this requires further model training on data from the target speaker, which is a costly process that may incorporate potentially sensitive data into server-side models. In contrast, this work focuses on the zero-shot approach which avoids finetuning altogether, and instead uses pretrained speaker verification embeddings together with a jointly trained reference encoder to encode utterance-level information that helps capture aspects such as speaker identity and prosody. Subjective listening tests find that both utterance embeddings and a reference encoder improve the continuity of speaker identity and prosody between the edited synthetic speech and unedited original recording in the zero-shot setting.

Accent-Robust Automatic Speech Recognition Using Supervised and Unsupervised Wav2vec Embeddings

Speech recognition models often obtain degraded performance when tested on speech with unseen accents. Domain-adversarial training (DAT) and multi-task learning (MTL) are two common approaches for building accent-robust ASR models. ASR models using accent embeddings is another approach for improving robustness to accents. In this study, we perform systematic comparisons of DAT and MTL approaches using a large volume of English accent corpus (4000 hours of US English speech and 1244 hours of 20 non-US-English accents speech). We explore embeddings trained under supervised and unsupervised settings: a separate embedding matrix trained using accent labels, and embeddings extracted from a fine-tuned wav2vec model. We find that our DAT model trained with supervised embeddings achieves the best performance overall and consistently provides benefits for all testing datasets, and our MTL model trained with wav2vec embeddings are helpful learning accent-invariant features and improving novel/unseen accents. We also illustrate that wav2vec embeddings have more advantages for building accent-robust ASR when no accent labels are available for training supervised embeddings.

On lattice-free boosted MMI training of HMM and CTC-based full-context ASR models

Hybrid automatic speech recognition (ASR) models are typically sequentially trained with CTC or LF-MMI criteria. However, they have vastly different legacies and are usually implemented in different frameworks. In this paper, by decoupling the concepts of modeling units and label topologies and building proper numerator/denominator graphs accordingly, we establish a generalized framework for hybrid acoustic modeling (AM). In this framework, we show that LF-MMI is a powerful training criterion applicable to both limited-context and full-context models, for wordpiece/mono-char/bi-char/chenone units, with both HMM/CTC topologies. From this framework, we propose three novel training schemes: chenone(ch)/wordpiece(wp)-CTC-bMMI, and wordpiece(wp)-HMM-bMMI with different advantages in training performance, decoding efficiency and decoding time-stamp accuracy. The advantages of different training schemes are evaluated comprehensively on Librispeech, and wp-CTC-bMMI and ch-CTC-bMMI are evaluated on two real world ASR tasks to show their effectiveness. Besides, we also show bi-char(bc) HMM-MMI models can serve as better alignment models than traditional non-neural GMM-HMMs.

Kaizen: Continuously improving teacher using Exponential Moving Average for semi-supervised speech recognition

In this paper, we introduce the Kaizen framework that uses a continuously improving teacher to generate pseudo-labels for semi-supervised training. The proposed approach uses a teacher model which is updated as the exponential moving average of the student model parameters. This 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 it for frame-level hybrid hidden Markov model - deep neural network (HMM-DNN) models and sequence-level connectionist temporal classification (CTC) based models. The proposed approach shows more than 10% word error rate (WER) reduction over standard teacher-student training and more than 50\% relative WER reduction over 10 hour supervised baseline when using large scale realistic unsupervised public videos in UK English and Italian languages.

Large scale weakly and semi-supervised learning for low-resource video ASR

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.

CHiME-6 Challenge:Tackling Multispeaker Speech Recognition for Unsegmented Recordings

Following the success of the 1st, 2nd, 3rd, 4th and 5th CHiME challenges we organize the 6th CHiME Speech Separation and Recognition Challenge (CHiME-6). The new challenge revisits the previous CHiME-5 challenge and further considers the problem of distant multi-microphone conversational speech diarization and recognition in everyday home environments. Speech material is the same as the previous CHiME-5 recordings except for accurate array synchronization. The material was elicited using a dinner party scenario with efforts taken to capture data that is representative of natural conversational speech. This paper provides a baseline description of the CHiME-6 challenge for both segmented multispeaker speech recognition (Track 1) and unsegmented multispeaker speech recognition (Track 2). Of note, Track 2 is the first challenge activity in the community to tackle an unsegmented multispeaker speech recognition scenario with a complete set of reproducible open source baselines providing speech enhancement, speaker diarization, and speech recognition modules.

Automatic Speech Recognition and Topic Identification for Almost-Zero-Resource Languages

Automatic speech recognition (ASR) systems often need to be developed for extremely low-resource languages to serve end-uses such as audio content categorization and search. While universal phone recognition is natural to consider when no transcribed speech is available to train an ASR system in a language, adapting universal phone models using very small amounts (minutes rather than hours) of transcribed speech also needs to be studied, particularly with state-of-the-art DNN-based acoustic models. The DARPA LORELEI program provides a framework for such very-low-resource ASR studies, and provides an extrinsic metric for evaluating ASR performance in a humanitarian assistance, disaster relief setting. This paper presents our Kaldi-based systems for the program, which employ a universal phone modeling approach to ASR, and describes recipes for very rapid adaptation of this universal ASR system. The results we obtain significantly outperform results obtained by many competing approaches on the NIST LoReHLT 2017 Evaluation datasets.