Release prep: Phase 3 LM (checkpoints/phase3/p3_lm_full), versions/voices/settings, auto-tagger, Gradio app, CC-BY-SA-4.0 weights + Apache-2.0 code, model card
85bc603 verified | # Python 3.10. Install torch/torchaudio for your CUDA first (Colab already has them), then: | |
| # pip install -r requirements.txt | |
| # and clone CosyVoice (see README). Versions below are the ones the samples were rendered with | |
| # (pins loosened where Colab's preinstalled versions also work). | |
| torch>=2.3 | |
| torchaudio>=2.3 | |
| transformers==4.51.3 | |
| huggingface_hub>=0.23 | |
| numpy==1.26.4 | |
| soundfile>=0.12 | |
| librosa>=0.10 | |
| onnxruntime-gpu>=1.18 | |
| HyperPyYAML>=1.2.2 | |
| modelscope>=1.20 | |
| conformer==0.3.2 | |
| diffusers==0.29.0 | |
| lightning>=2.2.4 | |
| omegaconf>=2.3.0 | |
| hydra-core==1.3.2 | |
| inflect>=7.3 | |
| x-transformers>=2.11 | |
| openai-whisper>=20231117 | |
| einops>=0.8 | |
| # pyworld>=0.3.4 # CosyVoice training only; infer.py runs without it | |
| pyarrow>=18 | |
| wget==3.2 | |
| gdown>=5.1 | |
| rich>=13.7 | |
| matplotlib>=3.7 | |
| protobuf>=4.25 | |
| pydantic>=2.7 | |
| wetext==0.0.4 | |
| audiotsm==0.1.2 # WSOLA time-stretch for --speed (falls back to librosa if missing) | |
| gradio>=4.44,<6 # app.py | |
| # optional, for --lid glotlid / the GlotLID checkbox: fasttext (or fasttext-wheel) | |