Blog: 微信文章
Lossless speculative decoding for speech models — trained drafters, serving recipes, and end-to-end benchmarks.
| Part | Doc | TL;DR |
|---|---|---|
| Speech LLM (ASR) | README_SPEECH_LLM.md | EAGLE3 / DFlash drafts for Qwen3-Omni-30B and Qwen2-Audio in sglang: 1.5–1.65x at batch=1, WER/CER unchanged. benchmark_speech_llm.py reproduces everything end to end. |
| TTS (text → speech tokens) | README_TTS.md | DSpark / DFlash drafters for CosyVoice2/3 LLMs in vLLM: 1.2–3.1x across batch sizes. benchmark_tts.py. |
| Training | training/README.md | How the drafts above were trained: SpecForge (speech LLM) and the yuekaizhang/speculators (tts branch) fork (TTS), both on target-regenerated (on-policy) labels. |
All draft checkpoints are on the HF hub under yuekai; each doc lists its exact model/dataset repos, framework versions, and measured numbers.