VoiceWeaver

Lightweight migration repo for the private CosyVoice / Fun-CosyVoice3 server setup.

This repository intentionally does not vendor the full CosyVoice source tree, conda env, pretrained weights, generated audio, or runtime logs. It stores only the small pieces needed to recreate the current server state:

• patches/cosyvoice-centos7-cosyvoice3.patch: local CosyVoice changes for CentOS 7 inference and Fun-CosyVoice3 WebUI compatibility.
• deploy/centos7/: private-server start/stop scripts and known-good requirements files.
• docs/: deployment notes and the streaming RL TTS design direction.
• scripts/bootstrap_cosyvoice.sh: rebuild helper that clones CosyVoice and applies the patch.

Current Server Snapshot

• Server workspace: /home/yanglh/L/tts_lab
• Working CosyVoice repo: /home/yanglh/L/tts_lab/CosyVoice
• Private conda env: /home/yanglh/L/tts_lab/.env
• WebUI bind: 127.0.0.1:50000
• Current model path inside CosyVoice: pretrained_models/Fun-CosyVoice3-0.5B
• CosyVoice base commit used for the patch: ace7c47f41bbd303aa6bf1ea80e6f9fbd595cd40

Rebuild Shape

bash scripts/bootstrap_cosyvoice.sh /path/to/workspace

Then create an environment suitable for the target machine and install one of:

python -m pip install -r deploy/centos7/requirements.cosyvoice-centos7-webui.txt

Download the model separately. On the lab network, ModelScope worked better than Hugging Face:

from modelscope import snapshot_download
snapshot_download(
    "FunAudioLLM/Fun-CosyVoice3-0.5B-2512",
    local_dir="pretrained_models/Fun-CosyVoice3-0.5B",
)

Start / Stop

The deployment scripts assume the private layout used on the lab server. If the workspace changes, edit BASE_DIR at the top of the scripts.

bash deploy/centos7/start_cosyvoice_webui.sh
bash deploy/centos7/stop_cosyvoice_webui.sh

The WebUI binds to localhost by default. Use an SSH tunnel from your local machine:

ssh -N -L 50000:127.0.0.1:50000 gpu2-server

Then open http://127.0.0.1:50000.