Released an open-source program for creating transcripts of meetings (for example, in Teams), calls, and lectures, with automatic screenshots. And an article about it (RUS/UKR/EN). Link in the comments.
It didn’t start with meetings. I wanted a tool that would listen as I commented aloud while testing an application, and then turn it, together with screenshots, into a document or nearly finished bug reports. But in the end, 9 out of 10 times I use it, it’s for call transcripts and as an assistant right during the conversation. And it’s also great at turning YouTube lectures into readable articles with slides.
Recognition is fully local. Call audio and screen screenshots never leave your machine. Three engines to choose from — faster-whisper, GigaAM (for Russian, 20 times faster than Whisper and without hallucinations like To be continued…”) and Parakeet for a mix of Russian and English.
What it can do:
— Text appears practically in real time, with timestamps, and a regular .txt file on disk is updated as it goes.
— Ctrl-Shift-S takes a screenshot, extracts the text from it via OCR, and inserts it directly into the right place in the transcript. There’s an automatic mode — it takes a screenshot every second, then a cascade of filters keeps only new slides. On a test 11-minute talk, 692 frames → 10 scenes, not a single slide lost.
— Call audio can be captured from the microphone or directly from an application (for example, Teams or Chrome/YouTube)
— Speaker diarization based on voice and context. If a person with voice A calls a person with voice B Vasya, then all of B’s lines in the resulting transcript will be labeled as Vasya. The Teams transcript itself can be loaded as a second source, so more accurate information about who said what can be taken from it.
— The Beautify button turns the raw transcript into a document: decisions, tasks, clean dialogue. While recording is in progress, you can ask the chat what else we’ve forgotten to discuss, for example.
The application definitely works on macOS (DMG in releases, macOS 14+, models are downloaded on first launch). On other platforms, a lot of it won’t work, only the basics.
