CosyVoice 3
Zero-shot multilingual voice cloning TTS across 9 languages and 18 Chinese dialects
About
CosyVoice 3 is the third generation of Alibaba's FunAudioLLM text-to-speech family, built for zero-shot voice cloning in the wild. The 0.5B release runs a three-stage pipeline: text tokenized with a Qwen2 BPE tokenizer feeds an autoregressive language model head that emits speech tokens, which a flow-matching diffusion transformer turns into mel spectrograms, rendered to 24 kHz audio by a causal HiFi-GAN vocoder. It supports streaming synthesis and ships under Apache 2.0, which makes it one of the few permissively licensed voice-cloning models. Quality varies by language and dialect, low-resource dialects trail behind, and like all open voice-cloning models it can be misused for impersonation, so deployments need consent guardrails.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: outperforms CosyVoice 2.0 on content consistency, speaker similarity, and prosody naturalness (paper arXiv 2505.17589)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: arXiv paper 2505.17589 (27 May 2025) and official demo page funaudiollm.github.io/cosyvoice3 verified this session; HF release FunAudioLLM/Fun-CosyVoice3-0.5B-2512 and Apache-2.0 license confirmed via community GGUF readme; Exact weights release day not verified — releasedAt null.