# Meeting notes on Linux, without a bot or a cloud account.

> Record mic and call audio on Linux with no meeting bot, transcribe on your computer and replay the moment behind each note. Open-source developer preview.

Source: https://overtalk.io/linux.html · Facts checked: 2026-10-20

Overtalk is an MIT-licensed meeting notebook that runs on Linux. It records your microphone and the call's audio through PulseAudio, including PipeWire's PulseAudio server, so no bot joins Zoom, Google Meet or Teams. A local speech engine transcribes the recording and a local model writes notes, each linked to the moment it was said. As of 20 October 2026 it is a developer preview built from source, with no installer yet.

**Try it without installing:** [the sample meeting](https://overtalk.io/index.html#sample) plays a fictional review and the notes Overtalk saved for it.

**Build it:** [the build guide](https://overtalk.io/setup.html) lists what you need.

## Meeting-notes tools that run on Linux

Most commercial note-takers have no Linux app. These do, or can be built for it. Each row was read on the project's own pages on 20 October 2026; "not checked" means we did not confirm it, not that it is missing.

Meeting-notes tools on Linux, as of 20 October 2026

| Tool | How you get it on Linux | Transcription | Notes | Licence |
| --- | --- | --- | --- | --- |
| Overtalk | Built from source; .deb and AppImage being prepared | On your computer: Parakeet built in, Qwen3-ASR with the optional speech worker | On your computer by default; or your own Privatemode account | MIT |
| [Meetily](https://meetily.ai/granola-for-linux) | Built from source; a native app is in development | On your computer (Whisper or Parakeet) | Ollama on your computer, or your own key for a cloud provider | MIT community edition; Pro is closed and not on Linux |
| [OpenWhispr](https://github.com/OpenWhispr/openwhispr) | AppImage, .deb, .rpm | On your computer (Whisper or Parakeet), or cloud | Not checked | MIT |
| [Anarlog](https://anarlog.so/download) | Beta: .deb, AppImage, APT, AUR | A transcription server you run, or a cloud provider; built-in local transcription needs Apple Silicon | Local models or your own key; Pro uses the cloud | MIT; enterprise parts licensed separately |
| [Mnemosyne](https://github.com/corpetty/mnemosyne) | .deb, .rpm, AppImage, Flatpak bundle | On your computer | On your computer (llama.cpp or Ollama), or cloud | MIT |
| [StenoAI](https://github.com/stenolabs/stenoai) | Alpha AppImage | On your computer (Parakeet or Whisper) | Ollama on your computer | MIT |
| [Minutes](https://github.com/silverstein/minutes) | Command line only on Linux | On your computer | On your computer, as Markdown | MIT |
| [Loomfy](https://flathub.org/apps/ai.loomfy.Loomfy) | Flathub | On your computer | Cloud, on paid plans | Closed source |

If you only need a transcript, Linux has good offline transcribers: [Vibe](https://github.com/thewh1teagle/vibe) records system audio and the microphone, [Buzz](https://github.com/chidiwilliams/buzz) and [Speech Note](https://github.com/mkiol/dsnote) transcribe files and dictation, and [aTrain](https://github.com/JuergenFleiss/aTrain) is built for research interviews. They make transcripts rather than meeting notes. For prices, bots and cloud services, see [the full comparison](https://overtalk.io/compare.html).

## Record a call on Linux: by hand, or with Overtalk

On Linux, the call's audio is available as the "monitor" source of your output device in PulseAudio or PipeWire. You can record it with ffmpeg and transcribe the file with whisper.cpp; you then have to deal with timing, speakers and notes yourself. Overtalk records both sources together, checks them with a sound check before you start, and transcribes and writes notes on your computer.

By hand, on PipeWire with pipewire-pulse or on PulseAudio:

`# Call audio and your microphone, mixed into one 16 kHz file ffmpeg -f pulse -i @DEFAULT_MONITOR@ -f pulse -i @DEFAULT_SOURCE@ \ -filter_complex amix=inputs=2:duration=shortest -ac 1 -ar 16000 call.wav # Transcribe it with whisper.cpp whisper-cli -m ggml-base.bin -f call.wav -otxt`

We checked that the ffmpeg command opens both devices on PipeWire 1.6 with pipewire-pulse on 20 October 2026. Press q to stop it. Tell everyone before you start: [what to say, and what the law asks](https://overtalk.io/recording-consent.html).

What Overtalk adds to that:

- A sound check of the microphone and the call audio before you record.

- Microphone and call audio mixed so neither drowns the other, kept in step, and saved for playback.

- A transcript you can correct, with speaker names you can edit.

- Notes in which every decision, action and open question plays the moment it came from.

- A recording notice to paste or read, saved with the meeting.

- Exports to PDF, DOCX, Markdown, text and JSON.

## How Overtalk captures audio on Linux

Overtalk talks to the desktop's PulseAudio-compatible server, which on most current distributions is PipeWire's pipewire-pulse. The microphone and call-audio menus use the same names as your sound settings. Choosing an output records that output's monitor, so whatever Zoom, a browser tab with Meet or Teams, or any other app plays through it is captured. Speakers and headphones on the same output share one monitor.

Overtalk never changes your default devices, profiles, volume or PipeWire's quantum. Each recording opens its own audio connection at 48 kHz and releases it when you stop or the sound server goes away.

## Processors, GPUs and models

- **Processor:** a source build targets your own processor. The coming packages need an x86-64-v3 processor (Intel Haswell or AMD Zen, 2013 onwards) and run on the CPU.

- **GPU:** optional. CUDA (NVIDIA) and Vulkan are build options when you build from source.

- **Models:** two downloads, chosen in the app, each pinned and checksummed: Parakeet TDT 0.6B v3 for speech (670 MB, 25 European languages) and a notes model of 1.0 to 2.6 GB. Starting the app downloads nothing.

- **More languages:** the optional local speech worker runs Qwen3-ASR 1.7B, with 30 languages including Arabic, Chinese, Hindi and Japanese, in its own Python environment.

## What is not ready yet

- No released package. A .deb for Ubuntu 24.04 or newer and Debian 13 or newer, and an AppImage, are built and being tested.

- No Flatpak. The model workers run in a bubblewrap sandbox, which needs a Flatpak-specific path first.

- macOS and Windows builds and their audio capture are not validated.

- The first run needs 1.7 to 3.3 GB of model downloads before the first transcript.

- No phone app and no shared team workspace.

## Questions about Overtalk on Linux

### Does Overtalk work with Zoom, Google Meet and Teams on Linux?

Yes. It records the audio your output device plays, so it works with the Zoom client and with Meet or Teams in a browser, without a bot joining the call. Choose the output your call plays through in the sound check.

### Do I need PipeWire?

No. Overtalk needs a PulseAudio-compatible sound server: PulseAudio itself, or PipeWire with pipewire-pulse, which most current distributions use.

### Is there a Flatpak, .deb or AppImage?

Not yet. A .deb and an AppImage are built and being tested for the first release; a Flatpak needs changes to how the model workers are sandboxed. Until then, [build from source](https://overtalk.io/setup.html).

### Does anything leave my computer?

Not with local models: the recording, transcript and notes are processed and stored on your computer, and no account is needed. Model downloads need a network connection. [Where a meeting goes on each route](https://overtalk.io/privacy.html).