A local-first desktop app for dictation. Switch between free local models and paid cloud providers in one click. Transcription types straight at your cursor — in any app.
/ what you get
Every feature built so you actually use it daily — not a demo reel.
Run faster-whisper on your own machine. No audio leaves your computer.
Eight model sizes from tiny (40MB) to large-v3 (3GB). Works offline.
No account. No cloud. No leak.
OpenAI gpt-4o-transcribe, Deepgram Nova-3, AssemblyAI Universal, OpenRouter, Speaches.
Bring your own API key. Switch in one click.
Local for drafts. Cloud for final takes.
German podcast, English code comments, mixed-language meeting notes — each profile has its own model, post-processing, and hotkeys. Switch on the fly.
One app. Every use case.
Every cloud call is logged with tokens, duration, and cost. Export to CSV. Set per-profile budgets. Never get surprised by a bill.
Receipts for every second.
/ workflow
Choose your language, model, and post-processing in one click. Presets for podcasting, code comments, email, and notes ship by default.
A floating indicator shows it's listening. Pause, think, keep going. Live partial results appear so you can course-correct mid-sentence.
Transcription is typed directly into whatever has focus — VS Code, Notion, your CMS, a terminal, an email draft. No copy-paste.
/ for developers
A local HTTP API so other apps can tap into your configured models. One endpoint, any ASR engine, unified response shape.
$ curl -X POST http://localhost:8123/transcribe \
-F "audio=@meeting.wav" \
-F "profile=meeting-notes"
> {
"text": "Kickoff on Monday, agenda attached...",
"profile": "meeting-notes",
"model": "faster-whisper/large-v3",
"duration_ms": 2480,
"cost_eur": 0.00
}/ what people are saying
"I dictate podcast scripts in German with local Whisper, then switch to the cloud profile for final polish. Twenty minutes saved per episode, every episode."
"Replaced three separate SDKs with one local API. Cost tracking alone earned it a permanent spot in my dotfiles."
"Teaching online means switching languages mid-session. Profiles make that trivial. Also: actually private. No upload."
"The floating indicator is what sold me. No guessing whether it's listening. And the open-source license means I can ship it to my team without procurement drama."
Used Smart Voice Flow? Send a short note — honest feedback is how this project grows.
Send a review/ get it
Two platforms at launch. macOS is on the way.
X11 and Wayland. Python 3.11+ recommended.
Signed binary. UAC-aware installer.
On macOS? Help us test so we can ship it sooner.
/ questions
gpt-4o-transcribe), Deepgram (Nova-3), AssemblyAI (Universal), OpenRouter,
and self-hosted Speaches.
large-v3 local is very close to cloud accuracy, especially for clean audio.
For noisy or niche-vocabulary audio, cloud models still win on average. You can A/B in the app.