CosyVoice 3 ONNX

ONNX builds of CosyVoice 3, made by Moonbox for on-device use with ONNX Runtime (CPU, CUDA, DirectML). Each precision is one folder holding the graphs and a small Python pipeline.py that does the rest (features, prompt, sampling loop); what the graphs take and return is under How to run.

Files

The last column is the benchmark on Moonbox's TTS set (voice cloning of English and Chinese sentences from a LibriSpeech and an AISHELL-1 prompt; details under How to run): error rate of a Whisper transcript against the text, speaker similarity to the prompt and DNSMOS, against the original PyTorch model.

Precision File Size sha256 Benchmark
fp16 fp16/campplus.onnx 28.3 MB a6ac6a639977 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/decoder.onnx 0.1 MB 1c9e47bfe992 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/decoder.onnx.data 720.0 MB b6fd909ffee6 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/f0_predictor.npz 26.5 MB 019d7d548514 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/flow_encoder.onnx 2.2 MB c8e9811f07e2 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/flow_estimator.onnx 663.8 MB d33b08c76f9c WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/flow_noise.npy 4.8 MB 1d0e8fbe9796 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/hift.onnx 329.4 MB 539b2813b28c WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/lm_head.onnx 12.1 MB df323aedb29d WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/mel80_24k.npy 0.3 MB 4f1aed49678e WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/pipeline.py 15.9 kB 4e50186a8f9a WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/speech_embed.onnx 12.1 MB 17d97744cef8 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/speech_tokenizer_v3.onnx 969.5 MB 23236a74175d WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/text_embed.onnx 272.3 MB 774a039af165 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/tokenizer.json 10.6 MB caa0dadcb8c4 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
fp16 fp16/whisper_mel128.npy 0.1 MB 364239f8b565 WER 0.0 % / CER 0.0 % (original 0.0 / 0.3, fp32), similarity 0.971 / 0.959 (0.973 / 0.964), DNSMOS 3.24 / 2.80 (3.22 / 2.84); RTF 0.53 (CUDA)
int8 int8/campplus.onnx 28.3 MB a6ac6a639977 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/decoder.onnx 0.2 MB a99aae751f62 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/decoder.onnx.data 384.5 MB d172d8771545 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/f0_predictor.npz 26.5 MB 019d7d548514 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/flow_encoder.onnx 2.7 MB db70cb6b9337 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/flow_estimator.onnx 334.7 MB 0c6385594675 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/flow_noise.npy 4.8 MB 1d0e8fbe9796 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/hift.onnx 329.4 MB 539b2813b28c WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/lm_head.onnx 6.1 MB b2a2a452902c WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/mel80_24k.npy 0.3 MB 4f1aed49678e WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/pipeline.py 15.9 kB 4e50186a8f9a WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/speech_embed.onnx 6.1 MB ec4fa58e761a WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/speech_tokenizer_v3.onnx 969.5 MB 23236a74175d WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/text_embed.onnx 136.7 MB f4adfc639d51 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/tokenizer.json 10.6 MB caa0dadcb8c4 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int8 int8/whisper_mel128.npy 0.1 MB 364239f8b565 WER 0.3 % / CER 0.0 %, similarity 0.969 / 0.955, DNSMOS 3.21 / 2.77; RTF 0.56 (CUDA)
int4 int4/campplus.onnx 28.3 MB a6ac6a639977 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/decoder.onnx 0.2 MB 7813d4b1a86e LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/decoder.onnx.data 205.6 MB 6904cf02d528 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/f0_predictor.npz 26.5 MB 019d7d548514 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/flow_encoder.onnx 2.7 MB db70cb6b9337 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/flow_estimator.onnx 334.7 MB 0c6385594675 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/flow_noise.npy 4.8 MB 1d0e8fbe9796 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/hift.onnx 329.4 MB 539b2813b28c LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/lm_head.onnx 6.1 MB b2a2a452902c LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/mel80_24k.npy 0.3 MB 4f1aed49678e LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/pipeline.py 15.9 kB 4e50186a8f9a LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/speech_embed.onnx 6.1 MB ec4fa58e761a LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/speech_tokenizer_v3.onnx 969.5 MB 23236a74175d LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/text_embed.onnx 136.7 MB f4adfc639d51 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/tokenizer.json 10.6 MB caa0dadcb8c4 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)
int4 int4/whisper_mel128.npy 0.1 MB 364239f8b565 LM 4-bit, DiT int8: WER 0.0 % / CER 0.0 %, similarity 0.971 / 0.957, DNSMOS 3.22 / 2.79; RTF 0.58 (CUDA)

How to run

CosyVoice 3 (Fun-CosyVoice3-0.5B-2512: FunAudioLLM's speech generator, voice cloning from a few seconds of reference audio and its transcript, Chinese, English and more, 24 kHz output) as ONNX graphs plus pipeline.py, a numpy runtime that reproduces the official CosyVoice3.inference_zero_shot(..., text_frontend=False) around them.

Graphs from the publisher, in every folder as published (Apache-2.0, the original repository's files): campplus.onnx (speaker embedding from 80-bin Kaldi fbank), speech_tokenizer_v3.onnx (speech tokens of the prompt from a 128-bin Whisper log-mel). Converted here from the original PyTorch modules:

  • decoder.onnx: the Qwen2-0.5B speech-token LM (onnxruntime-genai's model builder): inputs_embeds, attention_mask, past_key_values.{0..23}.key / .value [1, 2, past, 64] -> hidden_states [1, S, 896] and present.*; lm_head.onnx turns the last hidden state into the 6761 speech-token logits.
  • text_embed.onnx / speech_embed.onnx: the text and speech-token embeddings.
  • flow_encoder.onnx: speech tokens [1, N] and the speaker embedding [1, 192] -> the flow's condition mu [1, 80, 2N] and speaker vector spks [1, 80].
  • flow_estimator.onnx: the flow's 22-layer DiT, made from the publisher's flow.decoder.estimator.fp32.onnx. x, mu, cond [2, 80, T], mask [2, 1, T], t [2], spks [2, 80] -> velocity [2, 80, T].
  • hift.onnx: the HiFT vocoder with its neural source: mel [1, 80, T] and f0 [1, T] -> 24 kHz audio [1, 480 T]. Its STFT / inverse STFT are written as convolutions (ONNX has no inverse STFT), and its fixed noise is built in.

The f0 predictor runs in float64 in the original ("precision is crucial"), and ONNX Runtime has no float64 convolution: pipeline.py runs it in numpy, float64, from f0_predictor.npz. The flow's fixed starting noise is flow_noise.npy. pipeline.py also does the frontend in numpy (Kaldi fbank, Whisper log-mel, the Matcha mel of the prompt, and torchaudio's resampler, since another resampler changes the prompt mel noticeably), the LM loop with the original's repetition-aware nucleus sampling (top-p 0.8, top-k 25, window 10), the silent-token filter, the flow's 10 Euler steps with classifier-free guidance 0.7, and the vocoder. It needs numpy, onnxruntime (or onnxruntime-gpu), tokenizers and soundfile:

python fp16/pipeline.py "The quick brown fox jumps over the lazy dog." out.wav --ref voice.wav \
    --ref-text "Exactly what is said in voice.wav." --cuda --seed 0
import sys, soundfile as sf
sys.path.insert(0, "fp16")
import pipeline
tts = pipeline.CosyVoice3("fp16", "cuda")
ref, sr = sf.read("voice.wav", dtype="float32")
wav, out_sr = tts("Text to say.", ref, sr, "Transcript of voice.wav.", seed=0)       # 24 kHz

CosyVoice 3 needs <|endofprompt|> in its prompt text: the pipeline puts the official examples' "You are a helpful assistant.<|endofprompt|>" in front of the transcript. The official text normalisation (wetext) is not included; pass written-out text. Sampling uses numpy's generator, so a seed gives the same speech every run here, not the same speech as PyTorch with that seed. On CUDA the LM's KV cache stays on the GPU, and its session is created with enable_skip_layer_norm_strict_mode = 1; set it too if you run decoder.onnx yourself.

Precisions. fp16/: the LM in fp16; the embeddings, flow encoder and DiT estimator fp16 weights with fp32 math. int8/: the LM 8-bit MatMulNBits (block 32), the DiT and the small graphs weight-only int8, embedding tables int8 per row. int4/: the LM 4-bit MatMulNBits, the rest as int8/ (a 4-bit DiT put Chinese DNSMOS 0.24 below the original's). HiFT stays fp32 in every folder (83 MB). No bf16 build.

Benchmark. The lab's TTS set: 6 English sentences cloned from a LibriSpeech test-clean prompt and 4 Chinese sentences from an AISHELL-1 prompt, each with seeds 0-3 (40 clips per build), the original's sampling. Sampled speech can't be compared sample by sample, so each build is judged as the original is: Whisper large-v3-turbo's WER (English) or CER (Chinese) against the text, WavLM-base-plus-sv speaker similarity to the prompt, and DNSMOS (P.835 overall). The original is the official code (inference_zero_shot, text_frontend=False) in fp32 on CUDA. Pass (the TTS threshold used across these repositories): per language, error rate at most 2 points above the original's, similarity at most 0.03 below, DNSMOS at most 0.2 below. Real-time factor (processing time / audio time) on an RTX 3080 Laptop GPU, whole pipeline.

English WER similarity DNSMOS Chinese CER similarity DNSMOS Real-time factor
original (PyTorch fp32, CUDA) 0.0 % 0.973 3.22 0.3 % 0.964 2.84 1.5
fp16/ (CUDA) 0.0 % 0.971 3.24 0.0 % 0.959 2.80 0.53
int8/ (CUDA) 0.3 % 0.969 3.21 0.0 % 0.955 2.77 0.56
int4/ (CUDA) 0.0 % 0.971 3.22 0.0 % 0.957 2.79 0.58

Graphs checked against the original modules before the benchmark: HiFT 2.6e-5 from the generator's own inference, flow encoder 6e-6, embeddings and head 1e-6; the numpy frontend matches the original's features (speech tokens identical, speaker embedding 1e-5).

How these were made

From CosyVoice 3 at revision 29e01c4e8d. The graphs were exported from the original PyTorch modules with torch.onnx (opset 17 / 18) and checked against them; language-model decoders were built with onnxruntime-genai's model builder (inputs_embeds in, KV cache in and out). Weight-only quantization: int8 is a per-output-channel int8 weight behind DequantizeLinear (or 8-bit MatMulNBits in a text decoder), int4 is ONNX Runtime's MatMulNBits (block 32). Each card above says which precision every graph uses. Every precision was benchmarked against the original before upload (pass, per language: error rate at most 2 points above the original's, speaker similarity at most 0.03 below, DNSMOS at most 0.2 below), and the sha256 column above is what the files must hash to.

License

Apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this repository, and it governs your use of these files. The original model card follows unchanged.


Original model card

SVG Banners

👉🏻 CosyVoice 👈🏻

Fun-CosyVoice 3.0: Demos; Paper; Modelscope; Huggingface; CV3-Eval

CosyVoice 2.0: Demos; Paper; Modelscope; HuggingFace

CosyVoice 1.0: Demos; Paper; Modelscope; HuggingFace

Highlight🔥

Fun-CosyVoice 3.0 is an advanced text-to-speech (TTS) system based on large language models (LLM), surpassing its predecessor (CosyVoice 2.0) in content consistency, speaker similarity, and prosody naturalness. It is designed for zero-shot multilingual speech synthesis in the wild.

Key Features

  • Language Coverage: Covers 9 common languages (Chinese, English, Japanese, Korean, German, Spanish, French, Italian, Russian), 18+ Chinese dialects/accents (Guangdong, Minnan, Sichuan, Dongbei, Shan3xi, Shan1xi, Shanghai, Tianjin, Shandong, Ningxia, Gansu, etc.) and meanwhile supports both multi-lingual/cross-lingual zero-shot voice cloning.
  • Content Consistency & Naturalness: Achieves state-of-the-art performance in content consistency, speaker similarity, and prosody naturalness.
  • Pronunciation Inpainting: Supports pronunciation inpainting of Chinese Pinyin and English CMU phonemes, providing more controllability and thus suitable for production use.
  • Text Normalization: Supports reading of numbers, special symbols and various text formats without a traditional frontend module.
  • Bi-Streaming: Support both text-in streaming and audio-out streaming, and achieves latency as low as 150ms while maintaining high-quality audio output.
  • Instruct Support: Supports various instructions such as languages, dialects, emotions, speed, volume, etc.

Roadmap

  • 2025/12

    • release Fun-CosyVoice3-0.5B-2512 base model, rl model and its training/inference script
    • release Fun-CosyVoice3-0.5B modelscope gradio space
  • 2025/08

    • Thanks to the contribution from NVIDIA Yuekai Zhang, add triton trtllm runtime support and cosyvoice2 grpo training support
  • 2025/07

    • release Fun-CosyVoice 3.0 eval set
  • 2025/05

    • add CosyVoice2-0.5B vllm support
  • 2024/12

    • 25hz CosyVoice2-0.5B released
  • 2024/09

    • 25hz CosyVoice-300M base model
    • 25hz CosyVoice-300M voice conversion function
  • 2024/08

    • Repetition Aware Sampling(RAS) inference for llm stability
    • Streaming inference mode support, including kv cache and sdpa for rtf optimization
  • 2024/07

    • Flow matching training support
    • WeTextProcessing support when ttsfrd is not available
    • Fastapi server and client

Evaluation

Model Open-Source Model Size test-zh
CER (%) ↓
test-zh
Speaker Similarity (%) ↑
test-en
WER (%) ↓
test-en
Speaker Similarity (%) ↑
test-hard
CER (%) ↓
test-hard
Speaker Similarity (%) ↑
Human - - 1.26 75.5 2.14 73.4 - -
Seed-TTS ❌ - 1.12 79.6 2.25 76.2 7.59 77.6
MiniMax-Speech ❌ - 0.83 78.3 1.65 69.2 - -
F5-TTS ✅ 0.3B 1.52 74.1 2.00 64.7 8.67 71.3
Spark TTS ✅ 0.5B 1.2 66.0 1.98 57.3 - -
CosyVoice2 ✅ 0.5B 1.45 75.7 2.57 65.9 6.83 72.4
FireRedTTS2 ✅ 1.5B 1.14 73.2 1.95 66.5 - -
Index-TTS2 ✅ 1.5B 1.03 76.5 2.23 70.6 7.12 75.5
VibeVoice-1.5B ✅ 1.5B 1.16 74.4 3.04 68.9 - -
VibeVoice-Realtime ✅ 0.5B - - 2.05 63.3 - -
HiggsAudio-v2 ✅ 3B 1.50 74.0 2.44 67.7 - -
VoxCPM ✅ 0.5B 0.93 77.2 1.85 72.9 8.87 73.0
GLM-TTS ✅ 1.5B 1.03 76.1 - - - -
GLM-TTS RL ✅ 1.5B 0.89 76.4 - - - -
Fun-CosyVoice3-0.5B-2512 ✅ 0.5B 1.21 78.0 2.24 71.8 6.71 75.8
Fun-CosyVoice3-0.5B-2512_RL ✅ 0.5B 0.81 77.4 1.68 69.5 5.44 75.0

Install

Clone and install

  • Clone the repo

    git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
    # If you failed to clone the submodule due to network failures, please run the following command until success
    cd CosyVoice
    git submodule update --init --recursive
    
  • Install Conda: please see https://docs.conda.io/en/latest/miniconda.html

  • Create Conda env:

    conda create -n cosyvoice -y python=3.10
    conda activate cosyvoice
    pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com
    
    # If you encounter sox compatibility issues
    # ubuntu
    sudo apt-get install sox libsox-dev
    # centos
    sudo yum install sox sox-devel
    

Model download

from huggingface_hub import snapshot_download
snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')
snapshot_download('FunAudioLLM/CosyVoice-ttsfrd', local_dir='pretrained_models/CosyVoice-ttsfrd')

Optionally, you can unzip ttsfrd resource and install ttsfrd package for better text normalization performance.

Notice that this step is not necessary. If you do not install ttsfrd package, we will use wetext by default.

cd pretrained_models/CosyVoice-ttsfrd/
unzip resource.zip -d .
pip install ttsfrd_dependency-0.1-py3-none-any.whl
pip install ttsfrd-0.4.2-cp310-cp310-linux_x86_64.whl

Basic Usage

import sys
sys.path.append('third_party/Matcha-TTS')
from cosyvoice.cli.cosyvoice import AutoModel
import torchaudio

""" CosyVoice3 Usage, check https://funaudiollm.github.io/cosyvoice3/ for more details
"""
cosyvoice = AutoModel(model_dir='pretrained_models/Fun-CosyVoice3-0.5B')
# en zero_shot usage
for i, j in enumerate(cosyvoice.inference_zero_shot('CosyVoice is undergoing a comprehensive upgrade, providing more accurate, stable, faster, and better voice generation capabilities.', 'You are a helpful assistant.<|endofprompt|>希望你以后能够做的比我还好呦。',
                                                    './asset/zero_shot_prompt.wav', stream=False)):
    torchaudio.save('zero_shot_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)
# zh zero_shot usage
for i, j in enumerate(cosyvoice.inference_zero_shot('八百标兵奔北坡,北坡炮兵并排跑,炮兵怕把标兵碰,标兵怕碰炮兵炮。', 'You are a helpful assistant.<|endofprompt|>希望你以后能够做的比我还好呦。',
                                                    './asset/zero_shot_prompt.wav', stream=False)):
    torchaudio.save('zero_shot_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)

# fine grained control, for supported control, check cosyvoice/tokenizer/tokenizer.py#L280
for i, j in enumerate(cosyvoice.inference_cross_lingual('You are a helpful assistant.<|endofprompt|>[breath]因为他们那一辈人[breath]在乡里面住的要习惯一点,[breath]邻居都很活络,[breath]嗯,都很熟悉。[breath]',
                                                        './asset/zero_shot_prompt.wav', stream=False)):
    torchaudio.save('fine_grained_control_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)

# instruct usage, for supported control, check cosyvoice/utils/common.py#L28
for i, j in enumerate(cosyvoice.inference_instruct2('好少咯,一般系放嗰啲国庆啊,中秋嗰啲可能会咯。', 'You are a helpful assistant. 请用广东话表达。<|endofprompt|>',
                                                    './asset/zero_shot_prompt.wav', stream=False)):
    torchaudio.save('instruct_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)
for i, j in enumerate(cosyvoice.inference_instruct2('收到好友从远方寄来的生日礼物,那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐,笑容如花儿般绽放。', 'You are a helpful assistant. 请用尽可能快地语速说一句话。<|endofprompt|>',
                                                    './asset/zero_shot_prompt.wav', stream=False)):
    torchaudio.save('instruct_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)

# hotfix usage
for i, j in enumerate(cosyvoice.inference_zero_shot('高管也通过电话、短信、微信等方式对报道[j][ǐ]予好评。', 'You are a helpful assistant.<|endofprompt|>希望你以后能够做的比我还好呦。',
                                                    './asset/zero_shot_prompt.wav', stream=False)):
    torchaudio.save('hotfix_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)

Discussion & Communication

You can directly discuss on Github Issues.

You can also scan the QR code to join our official Dingding chat group.

Acknowledge

  1. We borrowed a lot of code from FunASR.
  2. We borrowed a lot of code from FunCodec.
  3. We borrowed a lot of code from Matcha-TTS.
  4. We borrowed a lot of code from AcademiCodec.
  5. We borrowed a lot of code from WeNet.

Citations

@article{du2024cosyvoice,
  title={Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens},
  author={Du, Zhihao and Chen, Qian and Zhang, Shiliang and Hu, Kai and Lu, Heng and Yang, Yexin and Hu, Hangrui and Zheng, Siqi and Gu, Yue and Ma, Ziyang and others},
  journal={arXiv preprint arXiv:2407.05407},
  year={2024}
}

@article{du2024cosyvoice,
  title={Cosyvoice 2: Scalable streaming speech synthesis with large language models},
  author={Du, Zhihao and Wang, Yuxuan and Chen, Qian and Shi, Xian and Lv, Xiang and Zhao, Tianyu and Gao, Zhifu and Yang, Yexin and Gao, Changfeng and Wang, Hui and others},
  journal={arXiv preprint arXiv:2412.10117},
  year={2024}
}

@article{du2025cosyvoice,
  title={CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training},
  author={Du, Zhihao and Gao, Changfeng and Wang, Yuxuan and Yu, Fan and Zhao, Tianyu and Wang, Hao and Lv, Xiang and Wang, Hui and Shi, Xian and An, Keyu and others},
  journal={arXiv preprint arXiv:2505.17589},
  year={2025}
}

@inproceedings{lyu2025build,
  title={Build LLM-Based Zero-Shot Streaming TTS System with Cosyvoice},
  author={Lyu, Xiang and Wang, Yuxuan and Zhao, Tianyu and Wang, Hao and Liu, Huadai and Du, Zhihao},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--2},
  year={2025},
  organization={IEEE}
}

Disclaimer

The content provided above is for academic purposes only and is intended to demonstrate technical capabilities. Some examples are sourced from the internet. If any content infringes on your rights, please contact us to request its removal.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for moonbox-dev/Fun-CosyVoice3-0.5B-2512-ONNX

Quantized
(32)
this model

Papers for moonbox-dev/Fun-CosyVoice3-0.5B-2512-ONNX