Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Voxtlm: unified decoder-only models for consolidating speech recognition/synthesis and speech/text continuation tasks

Soumi Maiti, Yifan Peng, Shukjae Choi, Jee-weon Jung, Xuankai Chang, Shinji Watanabe

TL;DR

VoxtLM introduces a decoder-only Voice-text Language Model that unifies speech recognition, speech synthesis, text generation, and speech continuation by merging text and discretized speech tokens into a single Voxt vocabulary. It uses a speech tokenizer and a speech token decoder to bridge continuous speech and discrete tokens, guided by special task tokens, and is trained with cross-entropy on publicly available data. The multitask setup yields strong improvements in TTS intelligibility and quality, and competitive gains in ASR and speech generation, with initialization from a pretrained text LM further enhancing performance. The work demonstrates reproducibility through open data, recipes, and model checkpoints, and points to the potential of fully unified speech-text generative models for scalable, multitask speech processing.

Abstract

We propose a decoder-only language model, VoxtLM, that can perform four tasks: speech recognition, speech synthesis, text generation, and speech continuation. VoxtLM integrates text vocabulary with discrete speech tokens from self-supervised speech features and uses special tokens to enable multitask learning. Compared to a single-task model, VoxtLM exhibits a significant improvement in speech synthesis, with improvements in both speech intelligibility from 28.9 to 5.6 and objective quality from 2.68 to 3.90. VoxtLM also improves speech generation and speech recognition performance over the single-task counterpart. Further, VoxtLM is trained with publicly available data and training recipes and model checkpoints are open-sourced to make fully reproducible work.

Voxtlm: unified decoder-only models for consolidating speech recognition/synthesis and speech/text continuation tasks

TL;DR

VoxtLM introduces a decoder-only Voice-text Language Model that unifies speech recognition, speech synthesis, text generation, and speech continuation by merging text and discretized speech tokens into a single Voxt vocabulary. It uses a speech tokenizer and a speech token decoder to bridge continuous speech and discrete tokens, guided by special task tokens, and is trained with cross-entropy on publicly available data. The multitask setup yields strong improvements in TTS intelligibility and quality, and competitive gains in ASR and speech generation, with initialization from a pretrained text LM further enhancing performance. The work demonstrates reproducibility through open data, recipes, and model checkpoints, and points to the potential of fully unified speech-text generative models for scalable, multitask speech processing.

Abstract

We propose a decoder-only language model, VoxtLM, that can perform four tasks: speech recognition, speech synthesis, text generation, and speech continuation. VoxtLM integrates text vocabulary with discrete speech tokens from self-supervised speech features and uses special tokens to enable multitask learning. Compared to a single-task model, VoxtLM exhibits a significant improvement in speech synthesis, with improvements in both speech intelligibility from 28.9 to 5.6 and objective quality from 2.68 to 3.90. VoxtLM also improves speech generation and speech recognition performance over the single-task counterpart. Further, VoxtLM is trained with publicly available data and training recipes and model checkpoints are open-sourced to make fully reproducible work.
Paper Structure (11 sections, 2 equations, 2 figures, 8 tables)

This paper contains 11 sections, 2 equations, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. VoxtLM
  5. Data format
  6. Training
  7. Inference
  8. Evaluation Metrics
  9. Experiments
  10. Results
  11. Conclusion

Figures (2)

  • Figure 1: ASR and TTS use encoder-decoder architecture while VoxtLM is decoder-only. In VoxtLM, all parameters are shared between speech and text modalities, compared to separate encoder/ decoder for speech and text.
  • Figure 2: Overview of VoxtLM, our proposed autoregressive decoder-only LM incorporating speech and text within an integrated vocabulary $\mathcal{V}_{\text{voxt}}$. The model uses two additional modules, the speech tokenizer and the speech token decoder to facilitate the conversion between continuous speech signal and discrete speech tokens.