TTS Channel: CosyVoice3

What is CosyVoice3

CosyVoice is an open-source TTS service from Alibaba's Tongyi Lab, supporting voice cloning in Chinese, English, Japanese, Korean, and Cantonese. CosyVoice3 is the latest version, with the following features:

• 5 languages supported: Chinese (zh), English (en), Japanese (jp), Korean (ko), Cantonese (yue)
• 3-second rapid cloning: clone a voice from just 3 seconds of reference audio
• Built-in clone character: directly replicate the voice from the original video

CosyVoice open-source: https://github.com/FunAudioLLM/CosyVoice

---

Prerequisites

Requirement	Details
pyVideoTrans version	Latest version recommended
Hardware	NVIDIA GPU recommended
Network	First launch downloads models from modelscope.cn

---

Important: WebUI File Modification

The official webui.py included with CosyVoice3 cannot be used directly! The audio component uses streaming output, which causes the API to return m3u8 files instead of wav audio.

Solution

Open webui.py, search for streaming=True and change it to streaming=False, then restart webui.py.

Alternatively, download a pre-modified webui.py to replace the official one:

• Download: https://github.com/jianchang512/stt/releases/download/0.0/cosyvoice3-webui-py.zip

---

Deployment

Option A: Windows Package (recommended for beginners)

• Baidu Cloud download: https://pan.baidu.com/s/1Lp1GNGZGaO6WQG3NLli_RQ?pwd=1234
• HuggingFace download: https://huggingface.co/mortimerme/repocollect/resolve/main/cosyvoice3-0529.7z?download=true

Download, extract, and double-click the startup script.

Option B: From Source

1. Deploy the official project
2. Confirm that webui.py starts successfully and you can perform a dubbing operation in the UI
3. Download the modified webui.py and replace the official one
4. Restart webui.py

---

Using in pyVideoTrans

Steps

1. Upgrade: Make sure pyVideoTrans is updated to the latest version
2. Start the WebUI: Ensure the CosyVoice project is deployed and webui.py is running. Open http://127.0.0.1:8000 in your browser
3. Configure address: Open pyVideoTrans → Menu → TTS Settings → CosyVoice, enter the WebUI address (default: http://127.0.0.1:8000)
4. Fill in reference audio: Enter the reference audio filename and corresponding text
5. Select channel: Choose CosyVoice from the "TTS Channel" dropdown on the main screen

Reference Audio Format

audio_filename.wav#text corresponding to the audio

Example: nverguo.wav#Queen of Women speaking

Place the audio file in the f5-tts folder under the pyVideoTrans project root directory.

Character Selection

• clone character: Replicates the voice from the original video (automatic cloning)
• custom character: Uses the configured reference audio for cloning

---

Reference Audio Requirements

Item	Requirement
Format	WAV (recommended)
Duration	3–10 seconds (ideal)
Content	Clear pronunciation, no background noise
Location	f5-tts folder under pyVideoTrans root directory

Source code verification (_cosyvoice.py): CosyVoice uses the /generate_audio API endpoint. Reference audio automatically gets an <|endofprompt|> tag added, and streaming output is forced off (stream: False).

---

Troubleshooting

1. Model Download Is Slow on First Use

The first launch automatically downloads models from modelscope.cn, which may take a while. Please be patient.

2. Returns m3u8 File Instead of Audio

This is because the WebUI file hasn't been modified. Search for streaming=True and change it to streaming=False, or download the modified webui.py.

3. Dubbing Quality Is Poor

• Ensure the reference audio is high quality (clear pronunciation, no noise)
• Reference audio duration should be 3–10 seconds
• Try different reference audio

4. Connection Failed

Make sure:

• CosyVoice WebUI is running
• The address is correct (default: http://127.0.0.1:8000)
• The browser can open the WebUI interface