Sovereign Voice: Your voice, on your silicon. No cloud required.

Hey Product Hunt, I built Sovereign Voice because I was paying a cloud voice service to store and synthesize my own voice, and one day it occurred to me: my MacBook has a Metal GPU and Apple Silicon compute sitting idle most of the time. Why is my voice living on a server I do not control? The problem is not that cloud voice services are bad. It is that their architecture is wrong for a certain kind of user. If you are a podcaster, a voiceover artist, or someone who records sensitive material — interviews, legal narration, confidential briefings — you have two options today: pay per character and trust a third party with your voice model, or use nothing. Sovereign Voice is a third option. The architecture is straightforward. You download the Mac app (v0.5.3). It installs a daemon that runs the Qwen3-TTS voice synthesis model via MLX on the Metal GPU. To synthesize, you open the studio window, type your script, click Synthesize. The audio is on your drive. Pull the ethernet cable. It still works. Latency, measured on the shipped DMG (n=30, library_neutral, 2026-09-05): cold p50 ~10s / p90 ~16s on an M4 Air; warm p50 ~4.6s / p90 ~5.5s. The first synthesis after install is a one-time ~113s — the model weights (~1.9 GB) copy to disk on first launch. I am publishing these numbers rather than saying "instant" because one request in ten takes ~16s cold, and hiding that would be dishonest. M0-G4 is contested — a re-bench showed a bimodal stall under memory pressure that I am investigating (see docs/gtm/FACTS.md §5). A few things I am proud of: - The full synthesis pipeline runs without an internet connection. - There is no account. Nothing to sign in to. - The voice model file lives on your disk. You own it. Back it up. Move it to a new Mac. - License keys are verified locally against an embedded Ed25519 public key — no runtime license server. - Optional anonymous usage ping, off by default — no audio, text, or personal data ever sent. You can verify with Little Snitch or `lsof -i`. I am not making a comparative voice-quality claim. The blind listening study (M0-G3) was withdrawn (decision 0025) — I could not recruit non-author listeners, and the claim was never worth the gate it was costing. The 45-second demo in the gallery was narrated by the app itself; listen and judge for yourself. The Mac app is free: a 25-generation soft cap (you can keep generating after 25 — no hard paywall) with one library voice and one free voice clone. Pro is $19 one-time and adds the other two bundled library voices (Warm, Crisp), additional clones, and long-form mode. That is the complete Pro set — three things, nothing else is sold. Drop your email after your second synthesis for +5 free generations — no purchase required. Happy to answer questions in the comments.

Sovereign Voice

Your voice, on your silicon. No cloud required.

Your voice, on your silicon. No cloud required.

Sovereign Voice

Previous Sovereign Voice Launches

Previous Sovereign Voice Launches