2025
Warble
Voice-to-text desktop app. Apple-signed, per-user vocab.
- Launched in
- 3 weeks
- Live for
- 6 months
On this page
Context
Where this lives in the world.
Warble is a native macOS voice-to-text app built for daily writing flow — capture thoughts as you speak, with offline-first recording and per-user vocabulary that learns the way you actually talk.
Challenge
What was hard about it.
Three things had to be true at once. Offline-first: recording cannot block on a network. Per-user: vocabulary and dataset paths are user-scoped from day one so future personalization is a config change, not a refactor. Native: feels like a Mac app, not a web wrapper.
Solution
What we built to solve it.
Tauri 2 for the shell — native menu bar integration, proper window behavior, sub-10MB binary. React 19 + Vite for the UI. Recording happens locally; transcription routes to a self-hosted Whisper instance with bearer-token auth. Three themes (Dawn / Ocean / Nebula) share the same design tokens but keep isolated palettes.
Tech story
Decisions worth calling out.
- Apple-signed and notarized for distribution outside the App Store — passes Gatekeeper without warnings
- Per-user config + dataset paths — scaffolding for future fine-tuning is already in place
- Whisper transcription pipeline routes through our own inference instead of OpenAI's API
- Three discrete themes with isolated token sets — Dawn / Ocean / Nebula don't share variables, so theme work is contained
- Offline-first recording — clipboard fallback when transcription endpoint is unreachable
Screenshots
What it looks like in flight.
Outcome
App Store candidate · daily-driver for the team
Looking to build something like this?
If you're sitting on a similar problem — multi-currency booking, clinical workflow, real-time pipelines, anything in this orbit — we've already done the hard parts once. The second time is faster, cheaper, and shaped to your business. Tell us what you have in mind.