Scribe

What does Scribe do?

Scribe describes itself as "Self-hosted, GPU-accelerated transcription with speaker labels. An open-source alternative to Otter.ai that runs entirely on your own hardware."

The pipeline is Whisper large-v3 for speech recognition and pyannote.audio's speaker-diarization-community-1 model for speaker labels. Audio goes in, a transcript with speaker labels comes out. The web UI is drag-and-drop, the REST API answers HTTP calls, and the CLI handles batch work. Export formats are plain text, SRT and VTT subtitles, JSON and Markdown.

Key facts

• Licence: MIT.
• Platforms: Windows and Linux with Python 3.12.
• Transcription model: OpenAI Whisper large-v3.
• Diarization model: pyannote/speaker-diarization-community-1, which needs a free Hugging Face token to download.
• Interfaces: drag-and-drop web UI, REST API, CLI.
• Export formats: TXT, SRT, VTT, JSON, Markdown.
• Hardware tested: NVIDIA RTX 3060 Ti with 8 GB of VRAM, per the README. CPU-only is supported but slow.
• Disk: about 6 GB for the models on first run.
• Install: clone the repository, create a Python 3.12 virtual environment, install requirements, start the server.

What does it replace, and where does it fall short?

Scribe replaces Otter.ai for teams that want to keep meeting audio on their own hardware and are happy to run a server with an NVIDIA GPU.

Where it falls short: no live meeting bot. Scribe takes an audio or video file; it does not join a Google Meet, Zoom or Microsoft Teams call. Otter.ai's selling point is live capture of a meeting, so Scribe only covers the transcription and diarization part of that flow. The README names Windows and Linux; macOS is not listed as a supported platform and Apple Silicon has no CUDA path. The diarization model needs a Hugging Face token, which adds a step for a first-time setup.

How does Scribe run?

Clone the repository, create a Python 3.12 virtual environment, install the requirements file, then run the server. On Linux it runs as a Python process; on Windows the same. Add a Hugging Face token so the pyannote model can download; the model's licence on Hugging Face requires the token on first fetch.

Open the web UI in a browser, drag an audio or video file in, and Scribe writes the transcript with speaker labels. The REST API is at the same host for programmatic calls, and the CLI handles batch work. Whisper large-v3 runs on the NVIDIA GPU where present and on the CPU otherwise.

Who is Scribe for?

A team or individual who records meetings or interviews and wants to transcribe them on their own hardware, with diarization. A reader who wants a bot that joins a Google Meet and transcribes live should look at Meetily or Echo. A single-speaker macOS use case is better served by Buzz or a desktop dictation app.

What limits does the README state?

• Windows and Linux only; macOS is not listed as a supported platform.
• Python 3.12 is required.
• An NVIDIA GPU with CUDA is recommended; the tested configuration is an RTX 3060 Ti with 8 GB of VRAM.
• CPU-only transcription is supported but slow.
• The diarization model needs a free Hugging Face token and accepts the model licence on first download.
• About 6 GB of disk for the models.

Questions people ask

Is Scribe open source?

Yes. The LICENSE file at the repository root is MIT, an OSI-approved licence. The Whisper weights are MIT by OpenAI. The pyannote speaker-diarization-community-1 model has its own licence on Hugging Face, which the user accepts when fetching the model with the Hugging Face token.

Does Scribe work offline?

Once the Whisper and pyannote models are downloaded, yes. Scribe runs as a self-hosted server on the user's own machine and the README says it runs "entirely on your own hardware." The only required network calls are the model downloads on first run.

Can Scribe be used commercially?

Yes for the code, which is MIT. Whisper's weights are MIT. The pyannote speaker-diarization-community model has its own terms, and a commercial user should read them on the model card before deploying. Scribe itself adds no extra restriction.

How does Scribe compare with Otter.ai?

Scribe does the transcription and diarization part of what Otter.ai sells, but without the bot that joins live meetings. It runs on the user's own hardware, keeps audio on that hardware, and is free. Otter.ai is a hosted product with live meeting capture, team workspaces and a mobile app. For live capture on the Mac, pair Scribe with Echo or Meetily. The open-source alternatives to Otter.ai page sets Scribe beside these.

What hardware does Scribe need?

The README names an NVIDIA RTX 3060 Ti with 8 GB of VRAM as the tested configuration. Any NVIDIA GPU with CUDA support is listed as compatible, and more VRAM speeds up Whisper large-v3 on long files. CPU-only transcription is supported but slow. About 6 GB of disk is needed for the Whisper and pyannote models on first run.

Sources

• Scribe README and LICENSE: github.com/jay739/scribe, read 2026-11-09.