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.
Recent progress in Automatic Speech Recognition (ASR) has been coupled with a substantial increase in the model sizes, which may now contain billions of parameters, leading to slow inferences even with adapted hardware. In this context, several ASR models exist in various sizes, with different inference costs leading to different performance levels. Based on the observation that smaller models perform optimally on large parts of testing corpora, we propose to train a decision module, that would allow, given an audio sample, to use the smallest sufficient model leading to a good transcription. We apply our approach to two Whisper models with different sizes. By keeping the decision process computationally efficient, we build a decision module that allows substantial computational savings with reduced performance drops.
This work introduces Zambezi Voice, an open-source multilingual speech resource for Zambian languages. It contains two collections of datasets: unlabelled audio recordings of radio news and talk shows programs (160 hours) and labelled data (over 80 hours) consisting of read speech recorded from text sourced from publicly available literature books. The dataset is created for speech recognition but can be extended to multilingual speech processing research for both supervised and unsupervised learning approaches. To our knowledge, this is the first multilingual speech dataset created for Zambian languages. We exploit pretraining and cross-lingual transfer learning by finetuning the Wav2Vec2.0 large-scale multilingual pre-trained model to build end-to-end (E2E) speech recognition models for our baseline models. The dataset is released publicly under a Creative Commons BY-NC-ND 4.0 license and can be accessed via https://github.com/unza-speech-lab/zambezi-voice .
Foreign accent conversion (FAC) is a special application of voice conversion (VC) which aims to convert the accented speech of a non-native speaker to a native-sounding speech with the same speaker identity. FAC is difficult since the native speech from the desired non-native speaker to be used as the training target is impossible to collect. In this work, we evaluate three recently proposed methods for ground-truth-free FAC, where all of them aim to harness the power of sequence-to-sequence (seq2seq) and non-parallel VC models to properly convert the accent and control the speaker identity. Our experimental evaluation results show that no single method was significantly better than the others in all evaluation axes, which is in contrast to conclusions drawn in previous studies. We also explain the effectiveness of these methods with the training input and output of the seq2seq model and examine the design choice of the non-parallel VC model, and show that intelligibility measures such as word error rates do not correlate well with subjective accentedness. Finally, our implementation is open-sourced to promote reproducible research and help future researchers improve upon the compared systems.
Electrolarynx is a commonly used assistive device to help patients with removed vocal cords regain their ability to speak. Although the electrolarynx can generate excitation signals like the vocal cords, the naturalness and intelligibility of electrolaryngeal (EL) speech are very different from those of natural (NL) speech. Many deep-learning-based models have been applied to electrolaryngeal speech voice conversion (ELVC) for converting EL speech to NL speech. In this study, we propose a multimodal voice conversion (VC) model that integrates acoustic and visual information into a unified network. We compared different pre-trained models as visual feature extractors and evaluated the effectiveness of these features in the ELVC task. The experimental results demonstrate that the proposed multimodal VC model outperforms single-modal models in both objective and subjective metrics, suggesting that the integration of visual information can significantly improve the quality of ELVC.
This paper presents a new database consisting of concurrent articulatory and acoustic speech data. The articulatory data correspond to ultrasound videos of the vocal tract dynamics, which allow the visualization of the tongue upper contour during the speech production process. Acoustic data is composed of 30 short sentences that were acquired by a directional cardioid microphone. This database includes data from 17 young subjects (8 male and 9 female) from the Santander region in Colombia, who reported not having any speech pathology.
Speech representations learned in a self-supervised fashion from massive unlabeled speech corpora have been adapted successfully toward several downstream tasks. However, such representations may be skewed toward canonical data characteristics of such corpora and perform poorly on atypical, non-native accented speaker populations. With the state-of-the-art HuBERT model as a baseline, we propose and investigate self-supervised adaptation of speech representations to such populations in a parameter-efficient way via training accent-specific residual adapters. We experiment with 4 accents and choose automatic speech recognition (ASR) as the downstream task of interest. We obtain strong word error rate reductions (WERR) over HuBERT-large for all 4 accents, with a mean WERR of 22.7% with accent-specific adapters and a mean WERR of 25.1% if the entire encoder is accent-adapted. While our experiments utilize HuBERT and ASR as the downstream task, our proposed approach is both model and task-agnostic.
In this paper, we introduce MFCCGAN as a novel speech synthesizer based on adversarial learning that adopts MFCCs as input and generates raw speech waveforms. Benefiting the GAN model capabilities, it produces speech with higher intelligibility than a rule-based MFCC-based speech synthesizer WORLD. We evaluated the model based on a popular intrusive objective speech intelligibility measure (STOI) and quality (NISQA score). Experimental results show that our proposed system outperforms Librosa MFCC- inversion (by an increase of about 26% up to 53% in STOI and 16% up to 78% in NISQA score) and a rise of about 10% in intelligibility and about 4% in naturalness in comparison with conventional rule-based vocoder WORLD that used in the CycleGAN-VC family. However, WORLD needs additional data like F0. Finally, using perceptual loss in discriminators based on STOI could improve the quality more. WebMUSHRA-based subjective tests also show the quality of the proposed approach.
Large language models (LLMs) have gained considerable attention for Artificial Intelligence Generated Content (AIGC), particularly with the emergence of ChatGPT. However, the direct adaptation of continuous speech to LLMs that process discrete tokens remains an unsolved challenge, hindering the application of LLMs for speech generation. The advanced speech LMs are in the corner, as that speech signals encapsulate a wealth of information, including speaker and emotion, beyond textual data alone. Prompt tuning has demonstrated notable gains in parameter efficiency and competitive performance on some speech classification tasks. However, the extent to which prompts can effectively elicit generation tasks from speech LMs remains an open question. In this paper, we present pioneering research that explores the application of prompt tuning to stimulate speech LMs for various generation tasks, within a unified framework called SpeechGen, with around 10M trainable parameters. The proposed unified framework holds great promise for efficiency and effectiveness, particularly with the imminent arrival of advanced speech LMs, which will significantly enhance the capabilities of the framework. The code and demos of SpeechGen will be available on the project website: \url{https://ga642381.github.io/SpeechPrompt/speechgen}
Self-supervised learning (SSL) based speech pre-training has attracted much attention for its capability of extracting rich representations learned from massive unlabeled data. On the other hand, the use of weakly-supervised data is less explored for speech pre-training. To fill this gap, we propose a weakly-supervised speech pre-training method based on speaker-aware speech data. It adopts a similar training procedure to the widely-used masked speech prediction based SSL framework, while incorporating additional target-speaker enrollment information as an auxiliary input. In this way, the learned representation is steered towards the target speaker even in the presence of highly overlapping interference, allowing potential applications to tasks such as target speech recognition. Our experiments on Libri2Mix and WSJ0-2mix datasets show that the proposed model achieves significantly better ASR performance compared to WavLM, the state-of-the-art SSL model with denoising capability.