requirements-cosyvoice2-topdeps.txt

torch==2.6.0

# Audio processing
numpy<2.0  # Must be <2.0 for ONNX compatibility
librosa>=0.10.0
soundfile>=0.12.0

# ONNX Runtime for encoder
onnxruntime==1.18.0

# Model loading and inference
modelscope>=1.20.0  # For downloading models from ModelScope
transformers==4.48.0  # For Qwen2 model in decoder # otherwise too new for NGC pytorch
diffusers>=0.35.0  # For Flow model in decoder

# CosyVoice specific dependencies
conformer==0.3.2  # Required by CosyVoice models
pyworld==0.3.4  # World vocoder for audio processing
wetext==0.0.4  # Text processing utilities
lightning>=2.2.4  # Required by CosyVoice models
hydra-core==1.3.2  # Config loading
omegaconf==2.3.0  # Config management
HyperPyYAML==1.2.2  # YAML config parsing
ruamel.yaml>=0.18.0  # YAML parsing for decoder config files

# Required utilities
einops  # Tensor operations
gdown>=5.1.0  # Model downloading
wget>=3.2  # File downloading
inflect>=7.3.0  # Text processing in decoder
kaldifst>=1.7.0  # FST operations for text processing
openai-whisper  # Whisper for audio transcription (used by decoder)
torchmetrics  # Metrics for model evaluation (used by decoder)
torch_complex
rotary_embedding_torch

# For running examples/notebooks
datasets==2.21.0  # For loading ESB diagnostic dataset (and for using ffmpeg instead of torchcodec)
matplotlib  # For visualization in notebooks
ipykernel  # For Jupyter kernel


# For running scipts/tokenizer-evaluation.py
static_ffmpeg
stempeg
pesq
pystoi
museval