I just shipped Sovereign Voice 0.5.2, a Mac app that runs voice synthesis entirely on-device using Qwen3-TTS through Apple's MLX framework on the Metal GPU. No account, no cloud round-trip, no API key. The model weights (~1.9 GB) download once on first launch and live on your disk — after that it works offline.
How it works
- Qwen3-TTS 0.6B (8-bit quantized, Apache-2.0 licensed)
- Inference via MLX on Apple Silicon Metal GPU
- Menu bar daemon + studio window
- First audio in ~1s; a full short sentence in ~5-6s on an M4
What you get for free
- The Mac app (signed + notarized DMG)
- 25 voice generations with one library voice profile
- Voice cloning from a short sample
- After 25 generations, the app keeps working — no hard paywall
Pro tier ($19 one-time)
- Unlimited generations
- All voice profiles + custom clones
- Studio mode (higher fidelity)
- Long-form synthesis
Privacy
- No audio, text, or personal data sent anywhere after the one-time weight download
- Telemetry is opt-in and anonymous (version, platform, synth count only)
- You can verify with Little Snitch or
lsof -i
Why a DMG and not the App Store
The App Store sandbox blocks the file I/O patterns the daemon needs for the local model and audio export.
Download
- Landing page: https://sovereign-voice-landing.pages.dev
- Direct DMG: https://sovereign-download.morning-lake-f944.workers.dev/releases/Sovereign-Voice-0.5.2-aarch64.dmg
I built this because I was paying a cloud voice service per character and wanted a fixed-cost, local-first tool for long-form work. Happy to answer questions about the architecture, MLX performance, or the model choice in the comments.
Sovereign Voice is voice AI you actually own — self-hosted, private by default. https://sovereign-voice-landing.pages.dev
Top comments (0)