Apple Silicon¶
Run OmniVoice natively on M-series Macs via PyTorch's MPS backend.
OmniVoice runs natively on Apple Silicon via PyTorch's MPS backend.
Install¶
Verify¶
You should see device=mps. The first call to any synthesis tool warms the model.
Force a Device¶
Or per-call:
Performance Tips¶
- First call is slow — checkpoint download + JIT warmup.
- Pre-warm with
omnivoice_load_model()before user-facing latency matters. - Lower
num_step(16) trades a little quality for ~2x speed. - Voice cloning with a 3-second reference is faster than 10s.
Known Issues¶
- The first model load on a fresh machine downloads ~1 GB. Subsequent runs reuse the HuggingFace cache (
~/.cache/huggingface/hub/). whisperASR (used byomnivoice_transcribeand auto-transcription) currently runs on CPU even when synthesis is on MPS — a small one-time cost.