TokenChain: A Discrete Speech Chain via Semantic Token Modeling

2025-10-07

AI Summary

This work addresses the bottleneck in joint ASR–TTS modeling within the speech chain by proposing the first fully discrete speech chain framework, using semantic tokens as a unified intermediate representation. Methodologically: (1) it introduces a two-stage discrete architecture—semantic-level ASR, autoregressive text-to-semantic modeling, and masked generative semantic-to-acoustic modeling; (2) it employs straight-through argmax and Gumbel-Softmax with dynamic temperature scheduling for stable training; (3) it establishes an end-to-end feedback mechanism to propagate cross-task information through the text interface and adopts dynamic weight averaging to enhance cross-domain transferability. Experiments show accelerated convergence by 2–6 epochs on LibriSpeech, with 5–13% relative WER reduction at equal training epochs; on TED-LIUM, ASR WER improves by 56% relatively and T2S WER by 31%, while catastrophic forgetting is significantly mitigated.

Technology Category

• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Machine Learning: Large Multimodal Models (LMMs)
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Machine Speech Chain, simulating the human perception-production loop, proves effective in jointly improving ASR and TTS. We propose TokenChain, a fully discrete speech chain coupling semantic-token ASR with a two-stage TTS: an autoregressive text-to-semantic model co-trained with ASR and a masked-generative semantic-to-acoustic model for synthesis only. End-to-end feedback across the text interface is enabled with straight-through argmax/Gumbel-Softmax and balanced with supervised ASR via dynamic weight averaging. Ablations examine optimal temperature schedules for in- and cross-domain transfer. Evaluation reveals TokenChain surpasses baseline accuracy 2-6 epochs earlier and yields 5-13% lower equal-epoch error with stable T2S on LibriSpeech, and reduces relative ASR WER by 56% and T2S WER by 31% on TED-LIUM with minimal forgetting, showing that chain learning remains effective with token interfaces and models.

Problem

• Proposes TokenChain to improve ASR and TTS through discrete semantic token modeling
• Enables end-to-end feedback via text interface with Gumbel-Softmax optimization
• Achieves faster convergence and lower error rates in speech recognition and synthesis

Innovation

• Discrete speech chain with semantic token modeling
• Two-stage TTS combining autoregressive and masked-generative models
• End-to-end feedback via straight-through gradient estimation