Privacy is a core constraint, not a feature toggle.
- All processing is local. Inference runs entirely on your Mac via MLX.
- No network calls except to download model weights from HuggingFace Hub on first use.
- No telemetry, no analytics, no cloud. Zero data leaves your machine after model download.
- Text stays on device. Input text is processed locally and never transmitted anywhere.
- Audio stays on device. Generated audio is written locally or played back on device.
| Boundary | Trust Level | Notes |
|---|---|---|
| User text input | Trusted | Processed locally, never transmitted |
| Generated audio | Trusted | Written to local filesystem or played on device |
| Model weights | Trusted | Downloaded from HuggingFace Hub (Apache 2.0, by hexgrad), cached locally |
| Voice style vectors | Trusted | Bundled with model weights, loaded from local cache |
No remote trust boundaries during inference. The only network access is the initial huggingface_hub.snapshot_download call to fetch model weights.
Report vulnerabilities responsibly:
- Do not open a public issue. Vulnerabilities stay private until a fix ships.
- Use GitHub's private vulnerability reporting to submit.
- Include:
- Steps to reproduce
- Demonstrated impact
- Suggested fix (if any)
Reports without reproduction steps or demonstrated impact are deprioritized.
Expect acknowledgment within 48 hours.
These are not considered vulnerabilities:
- Issues requiring physical access to the machine
- Adversarial text inputs designed to produce specific audio outputs
- Mispronunciations or inaccurate prosody
| Version | Supported |
|---|---|
| 0.1.x | Yes |