Skip to content

[New Model] CosyVoice 2/3 (LLM-based streaming TTS) #1552

Description

@lishunyang12

Model Information

Why support this model?

CosyVoice is the most popular open-source TTS project on GitHub with ~19k stars. CosyVoice 2/3 uses an LLM backbone (Qwen2.5-0.5B) for speech token generation combined with chunk-aware causal flow matching for waveform synthesis. This LLM-based architecture means vllm-omni's existing optimizations for transformer inference could directly benefit serving performance. CosyVoice 3 supports 9 languages and 18+ Chinese dialects.

Technical Notes

  • Two-stage: LLM (Qwen2.5-0.5B backbone) generates speech tokens via FSQ, then causal flow matching generates audio
  • Finite Scalar Quantization (FSQ) replaces VQ for full codebook utilization
  • Unified streaming/non-streaming synthesis (150ms latency in streaming mode)
  • CosyVoice 3 released mid-2025 with improved content consistency and prosody
  • Zero-shot cross-lingual voice cloning, emotion/dialect control
  • NVIDIA TensorRT-LLM runtime support contributed upstream

Closest supported model in vllm-omni

Qwen3-TTS / MiMo-Audio — CosyVoice shares the same LLM-based TTS architecture pattern: an LLM generates discrete speech tokens, then a vocoder/flow-matching module converts them to audio. Qwen3-TTS uses a similar Qwen backbone + Code2Wav stage, so much of the existing pipeline infrastructure (LLM inference, token-to-waveform conversion) could be reused.

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions