Research map / Large language models / Language, safety and interpretability
What Makes a Good Speech Tokenizer for LLM-Centric Speech Generation? A Systematic Study
Xiaoran Fan, Zhichao Sun, Yangfan Gao, Jingfei Xiong, Hang Yan, Yifei Cao, Jiajun Sun, Shuo Li, Zhihao Zhang, Zhiheng Xi, Yuhao Zhou, Senjie Jin and 10 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a decoupled tokenizer (FACodec) instead of coupled tokenizers inside the SLM (ablation) | LLM-centric SLM · LibriTTS test-clean | WER: 13.63 -> 6.07 | Improved |
| use decoupled FACodec instead of coupled BigCodec despite BigCodec's better reconstruction (ablation) | LLM-centric SLM · RoleTriviaQA in-domain | SIM (all speakers): 0.21 -> 0.61 | Improved |
| use semi-decoupled SpeechTokenizer instead of coupled tokenizers (ablation) | LLM-centric SLM · LibriTTS test-clean | UTMOS: 8 codebooks: 4.05 | Improved |
| add multi-token prediction with 12 heads to decoupled SLM (ablation) | FACodec SLM · LibriTTS test-clean | WER: 6.07 -> 3.01 | Improved |
| increase MTP compression from 3 to 12 tokens per hidden state (ablation) | FACodec SLM · LibriTTS test-clean | WER: 5.84 -> 3.01 | Improved |
| increase MTP compression from 1 to 12 tokens per hidden state (ablation) | FACodec SLM · LibriTTS test-clean | UTMOS: 3.93 -> 3.67 | Hurt |
| increase MTP compression from 1 to 12 tokens per hidden state (ablation) | FACodec SLM · LibriTTS test-clean | speech synthesis success rate: 0.87 -> 1.00 | Improved |
| increase MTP compression from 1 to 12 tokens per hidden state (ablation) | FACodec SLM · LibriTTS test-clean | SIM (seen speakers): 0.50 -> 0.47 | Hurt |
| add speaker embedding to NTP decoupled SLM (ablation) | FACodec NTP SLM · LibriTTS test-clean | WER: 6.07 -> 5.51 | Improved |
| add speaker embedding to MTP-3H decoupled SLM (ablation) | FACodec MTP-3H SLM · LibriTTS test-clean | WER: 5.84 -> 7.04 | Hurt |
| add speaker embedding to MTP-12H decoupled SLM (ablation) | FACodec MTP-12H SLM · LibriTTS test-clean | SIM: 0.47 -> 0.60 | Improved |
| add speaker embedding to MTP-12H decoupled SLM (ablation) | FACodec MTP-12H SLM · LibriTTS test-clean | WER: 3.01 -> 2.94 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- WavTokenizer — WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models (ICLR 2024 · similarity 0.88)
- A Unified Audio Language Model with Text-Aligned Factorized Audio Tokenization (NeurIPS 2026 · similarity 0.87)
- Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling (ICML 2025 · similarity 0.86)
- NTPP: Generative Speech Language Modeling for Dual-Channel Spoken Dialogue via Next-Token-Pair Prediction (ICML 2025 · similarity 0.86)
- TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling (NeurIPS 2025 · similarity 0.86)
- Enhancing Large Language Models through Adaptive Tokenizers (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
