Views
No views yet
| File | Size | Description |
|---|---|---|
audio_vae_encoder.onnx + .data | ~185 MB | Audio waveform to latent features |
audio_vae_decoder.onnx + .data | ~176 MB | Latent features to 48 kHz waveform |
voxcpm2_prefill.onnx + .data | ~7.8 GB | Text + reference audio to KV cache + DiT hidden state |
voxcpm2_decode_step.onnx + .data | ~8.1 GB | Single autoregressive decode step (10 CFM steps baked in) |
external_data=True — keep .onnx and .onnx.data files together.1# Install
2pip install voxcpm>=2.0.2 torch>=2.4.0 onnxruntime>=1.18.0 soundfile numpy tqdm huggingface_hub
3
4# Download models
5python -c "
6from huggingface_hub import snapshot_download
7snapshot_download('ai4all8/VoxCPM2-ONNX', local_dir='./onnx_models', ignore_patterns=['*.md', '*.txt'])
8"
9
10# Download VoxCPM2 PyTorch weights (needed for preprocessing)
11python -c "from voxcpm import VoxCPM; VoxCPM.from_pretrained('openbmb/VoxCPM2')"
12
13# Run inference (see GitHub repo for full CLI)
14git clone https://github.com/ai4all8/VoxCPM2-ONNX.git
15cd VoxCPM2-ONNX
16python infer.py --text "Hello!" --ref_wav speaker.wav --ref_text "Reference transcript."| Platform | RTF | Notes |
|---|---|---|
| AMD Ryzen 9 (Windows) | ~4.5x | 8 cores, ORT sequential |
| Intel Core (Linux) | ~9.5x | Single-threaded |