Your M4 MacBook Pro (32 GB RAM) is more than enough for Kokoro local.

Recommended plan: start with Kokoro ONNX (lowest infrastructure impact), then optionally test full PyTorch Kokoro.

1. Choose runtime

- Primary (recommended): kokoro-onnx
  - Officially claims near-real-time on M1 and lightweight model sizes.
  - Best for always-on CLI narration with minimal load.
- Secondary: kokoro (PyTorch)
  - More "native" pipeline, but heavier runtime.
  - On Apple Silicon, project docs suggest PYTORCH_ENABLE_MPS_FALLBACK=1.
