Verbale: meeting notes that never leave your machine

Who is speaking, without diarization

This is the idea the whole thing is built on, and it is why the transcripts are readable instead of a wall of unattributed text.

Telling speakers apart usually means diarization: a second model that clusters voices by timbre, gets confused by cross-talk, and changes its mind about how many people are in the room.

Verbale does none of that. Every audio output on PipeWire exposes a .monitor source carrying exactly what is being played, so Verbale records two tracks at once: your microphone, and your speakers. Your voice is only ever on one of them. Speaker attribution is not inferred, it is a property of how the audio was captured.

Both tracks come from one ffmpeg process, so they start together. Two processes drift apart by however long the second took to launch, and a transcript whose speakers are half a second out of order reads like nonsense.

The honest limitation: Them is one label for everyone on the far end. For minutes, "me or them" is usually the distinction that matters, and being right about that beats guessing wrong about four voices.

What it actually does

Four commands cover almost everything.

⏺

record

Captures both tracks, then transcribes and summarises when you stop. One command from silence to minutes.

☰

list & show

Every meeting you have recorded, and the notes for any of them, in the terminal.

⌕

web

A local page on loopback. Click a timestamp in the transcript and the recording jumps there.

↻

import

Already have a recording? Point Verbale at any audio file and get the same treatment.

Privacy you can verify, not trust

Any tool can claim to be private. This one is small enough that you can check.

One network call exists in the codebase, to 127.0.0.1:11434, your own Ollama. It is in summarize.py and takes two minutes to read.
Whisper runs on your CPU through faster-whisper. The audio never goes anywhere.
The weights download once from Hugging Face, deliberately, with verbale pull. After that you can stay offline forever.
The viewer binds to loopback. Pass a different host and it warns you in the terminal.
No account, no API key, no telemetry, no crash reporter, no subscription and no code path that could add one without it showing up in the diff.

Install

Linux with PipeWire or PulseAudio, and ffmpeg.

# the tool
pipx install git+https://github.com/MarcoZorn/verbale

# fetch the whisper weights once, then check everything
verbale pull
verbale doctor

# optional, for the summaries: any model you like
ollama pull qwen3:8b

# record a meeting. ctrl-c stops it.
verbale record "Sprint planning"
verbale show

Without Ollama you still get full transcripts; only the minutes are skipped. Recording is Linux only, because the capture layer is PulseAudio and PipeWire. verbale import works anywhere Python and ffmpeg do.

Your notes are just files

One directory per meeting. Nothing here needs Verbale to read it.

~/.local/share/verbale/2026-09-12-1030-sprint-planning/
├── meeting.json      what it is, how long, which models touched it
├── them.wav          what your speakers played
├── you.wav           what your microphone heard
├── transcript.json   segments with timings and speakers
├── transcript.md     the same thing, readable
└── notes.md          the minutes

A notes tool you cannot get your notes out of is a liability. This one is cat.