What Makes a Good Speech Tokenizer for LLM-Centric Speech Generation? A Systematic Study
This work addresses challenges in speech generation for AI systems, offering incremental improvements in tokenizer design and efficiency.
The study tackled the problem of designing effective speech tokenizers for speech-language models to improve cross-modal alignment and speech generation quality, finding that decoupled tokenization and multi-token prediction led to up to 12x faster decoding and reduced word error rate from 6.07 to 3.01.
Speech-language models (SLMs) offer a promising path toward unifying speech and text understanding and generation. However, challenges remain in achieving effective cross-modal alignment and high-quality speech generation. In this work, we systematically investigate the role of speech tokenizer designs in LLM-centric SLMs, augmented by speech heads and speaker modeling. We compare coupled, semi-decoupled, and fully decoupled speech tokenizers under a fair SLM framework and find that decoupled tokenization significantly improves alignment and synthesis quality. To address the information density mismatch between speech and text, we introduce multi-token prediction (MTP) into SLMs, enabling each hidden state to decode multiple speech tokens. This leads to up to 12$\times$ faster decoding and a substantial drop in word error rate (from 6.07 to 3.01). Furthermore, we propose a speaker-aware generation paradigm and introduce RoleTriviaQA, a large-scale role-playing knowledge QA benchmark with diverse speaker identities. Experiments demonstrate that our methods enhance both knowledge understanding and speaker consistency.