AllUSay
A private dictation app for the Mac. Speech recognition runs entirely on the device through Tether's QVAC runtime, and clean text is typed straight into whatever app you are using, with no account, no cloud and no tracking.
- Service
- macOS Application Design & Development
- Field
- Private dictation
- Engagement
- Ongoing (public beta)
The challenge
Deliver accurate, low-latency speech-to-text with no network at all: a fine-tuned model running on Apple Silicon, spoken corrections and automatic language detection, all inside a 130 MB app that respects the user's privacy by design.
What we built
- Native macOS app for Apple Silicon with push-to-talk and hands-free dictation, typing into any app at the cursor
- On-device speech recognition on QVAC, using a 250M-parameter model fine-tuned from NVIDIA Parakeet, downloaded once and run offline
- Automatic language detection across eight European languages, with spoken corrections in English, German and French
- Filler-sound removal, tone matching to the target app, and a personal dictionary and snippets with import and export
- Privacy-first product design: no sign-up, no analytics, and dictation history that stays on the device for as long as the user chooses
The outcome
Shipped as a public beta with a 4 to 8 percent word error rate on public datasets, about 150 ms latency for a sentence and a 15-minute dictation transcribed in about 12 seconds, all without a byte leaving the Mac.
