COSYVOICE / STREAMING TOKEN + FLOW TTS / 2025

Fun-CosyVoice 3 · 0.5B

TOTAL0.5Bparameters

ACTIVE0.5Bruntime path

DEPTHLLM + Flowrepeated blocks

HIDDEN≈1,024main trunk

CONTEXTstreamingstream / sequence

CHECKPOINT≈2.0GB system4 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

128 tokens → 10 seconds · 1.33T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT128TOKENS

TRANSFORMER STACKTOKEN LLM → FLOW → VOCODERdiscrete planning followed by continuous acoustic rendering

Three-Stage Residual

OUTPUT10SECONDS

wraps every repeated block

CAUSAL SPEECH-TOKEN LLM · TOKEN–FLOW CASCADE离散 LLM 与连续 Flow 分工，token 可边生成边交给声学模块。

Flow Acoustic Model连续生成阶段补足离散 token 没有携带的声学细节。

LLM≈0.5Bspeech tokens

FLOWconditionalacoustics

VOCODERneuralwaveform

MODEstreamingonline

VOICEzero-shotprompt

LICENSEApache-2.0code/weights