ling0322/libwaifu-cosyvoice-3

Fun-CosyVoice3-0.5B-2512 — the libwaifu package

Fun-CosyVoice3-0.5B-2512, converted to the package format libwaifu reads. A few seconds of somebody speaking and a sentence in; the sentence, in that voice, out, at 24 kHz. The same five models drawing the same audio; a different file layout and one table instead of five files.

This is a modified copy under the original Apache License 2.0 — see LICENSE and NOTICE for what was changed. It is not an official FunAudioLLM product and is not endorsed by them. The original weights are at FunAudioLLM/Fun-CosyVoice3-0.5B-2512.

What is here

package	on disk	what it is
cosyvoice3.yaml and its parts	4.4 GB	float32 throughout, 1 109 M parameters

One manifest carries the whole pipeline: the Qwen2-0.5B speech-token language model, the DiT flow, the HiFT vocoder, S3Tokenizer v3 and CAMPPlus, plus the tokenizer. Nothing else has to be fetched to speak.

Speaking with it

waifu -m cosyvoice                         # the page's text2speech tab
cargo run --release --example speak -- cosyvoice3.yaml voice.wav "你好。" out.wav

let tts = CosyVoice3::from_manifest(Device::Cuda, Residency::Device, &manifest)?;
let voice = tts.listen(&recording)?;                               // once per voice
let sound = tts.say("今天天气很好。", &voice, &options, &mut report)?;  // no transcript
let sound = tts.say_after("今天天气很好。", "录音里说的话。", &voice, &options, &mut report)?;

docs/cosyvoice3.md in the libwaifu repository has the full pipeline and how it was checked against upstream stage by stage.

Where this package knowingly differs from upstream

• Streaming, instruct mode (inference_instruct2), voice conversion and fp16 are not implemented.
• HiFT's random buffers are drawn from the reading's seed, and the flow's starting noise is upstream's own draw, stored as cosyvoice3.flow.rand_noise.

Attribution

Fun-CosyVoice3-0.5B-2512 is licensed under the Apache License 2.0 — see https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512. This package also bundles CAMPPlus from funasr/campplus (Apache License 2.0), converted the same way. See NOTICE for exactly what was changed.