Transcribe App Just Grew Up: Meeting Recorder, AI Action Items, and Mistral Voxtral

2026-06-02 · By Aram Adamyan

Tags: Products, AI Tools, Armenian AI, Productivity

Transcribe App started as a small itch. I wanted to press a key, speak, and have text appear, in Armenian, anywhere on my machine. No web app, no copy-paste from a browser, no Latin transliteration that I had to fix by hand.

A weekend project and 47 releases later, that small itch turned into a full-blown productivity tool. Here is what shipped in the latest run, and why every piece is there.

Press a hotkey, anywhere, in any app

The core promise has not moved. Hit Alt + R on Windows or Option + R on Mac, speak, hit Enter. The transcription is pasted right where your cursor is. Slack, Notion, Gmail, VS Code, the terminal, your Telegram DM. The app does not care which window has focus, it just types.

What changed: the hotkey is now fully rebindable, including mouse buttons, and it reads physical key codes so a French, German, or Armenian keyboard layout works without a workaround.

Meeting recorder with AI notes

This is the big one. Right-click the tray icon, hit Record Meeting, and Transcribe captures the call. You get back four things on disk:

A full transcript
A summary written in third person
A list of key decisions
Action items as checkboxes, with owner attribution when you provide an attendee list

The notes structure is borrowed from a Microsoft Research paper on meeting summarisation. It works. Phrases like "I will draft the spec" and "Aram should review it" get routed to the right person automatically. Long silence gaps are tagged as [speaker change] so the LLM can attribute who said what without needing a heavy diarization model.

The recorder pins itself on top with a red REC banner, a pulsing dot, a live timer, and an audio-level meter. You will not forget it is on.

Smart Meeting Mode: capture the other side

The classic pain of meeting recording is capturing the other participants, not just your own mic. On Windows, Transcribe now uses a dynamic WASAPI loopback mixer that scans for the active playback device and records computer sound and your microphone together. No virtual cables, no VB-Audio, no driver tinkering.

On macOS it routes through BlackHole, the free loopback driver. Install it once and you are set.

Three backends, your call

Transcribe ships with three transcription engines and a clean way to pick between them in the Models tab:

Local Whisper (tiny → large-v3). Fully offline, free forever, no API key, no data leaves your machine. The app grays out models your RAM cannot run.
Google Cloud Speech-to-Text. Best Armenian accuracy when you have internet and want enterprise quality.
Mistral Voxtral (Mini, Small, Large). New in this release, a strong cloud STT option for builders already inside the Mistral ecosystem.

For the AI summaries on meetings, you get the same flexibility: rule-based and free, local Qwen 1.5B/3B/7B for offline AI, or cloud models like Gemini 2.5 Flash, Claude 4.6/4.8, and GPT-5.4/5.5 with real-time pricing shown right in settings.

A real settings panel

Settings used to be a single dumping ground. It is now split into clean task-oriented tabs:

General for hotkey, default language, custom vocabulary, meeting recording mode, and privacy mode
Models with first-class cards for every local and cloud backend, plus a single integrated frame for API credentials
Staged transactional saves with Save and Cancel, so you can experiment without committing

Custom vocabulary is the underrated feature here. Seed it with your team names, product names, and Armenian transliterations of borrowed terms, and accuracy on your jargon jumps noticeably.

Auto-updates and a real installer

On Windows, the installer wizard handles Start Menu shortcuts, optional autostart, and optional desktop icon. When a new release ships, Transcribe notifies you and adds an Install update item to the tray menu. One click, no re-download dance.

On macOS, there is now a proper signed-looking .dmg. You still have to clear Gatekeeper once because the app is not notarised (the joys of being open source and not paying Apple $99 a year), but after that it just works.

Built for Armenian, useful for everyone

Transcribe was built first for Armenian speakers because that is the gap I felt every day. Native script, not Latin. Auto-detect that does not silently fall back to English. Cloud backends that actually handle Armenian at production quality.

But the same plumbing works for English, Russian, French, German, Spanish, and Arabic. If you want a hotkey-driven dictation tool that respects your privacy by default and grows up to handle full meeting workflows when you need it, this is for you.

Get it

Both builds are free, MIT-spirit non-commercial license, open source on GitHub.

If you find a bug, open an issue. If you build something interesting on top of it, tell me, I want to see it.