Skip to content

Repository files navigation

SpeechSpec

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.

About

Speculative Decoding for ASR and TTS

Resources

Stars

27 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages