scribe

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

No cloud, no accounts, no per-minute billing. Your audio never leaves the machine.

Features

Feature	Detail
Transcription	Whisper large-v3 via faster-whisper (CTranslate2, CUDA, int8_float16)
Speaker diarization	pyannote.audio, optional; without it scribe still transcribes and reports why labels are missing
Web UI	Drag and drop a file, watch progress, read a color-coded transcript
API	Upload, poll, fetch results as JSON
CLI	scribe transcribe meeting.mp3 --diarize -f srt
Exports	TXT, SRT, VTT, JSON, Markdown

Requirements

• NVIDIA GPU with CUDA support (tested on an RTX 3060 Ti, 8 GB VRAM). CPU works but is slow.
• Windows or Linux, Python 3.12 (the setup script installs its own via uv)
• Roughly 6 GB of disk for models on first run

Install

Quick start (Windows)

git clone https://github.com/jay739/scribe && cd scribe
powershell -ExecutionPolicy Bypass -File scripts/setup.ps1
.venv\Scripts\python -m scribe.server

Then open http://localhost:8323.

Speaker diarization setup

Diarization uses a gated Hugging Face model, so it needs a one-time token setup:

1. Create a free token at https://huggingface.co/settings/tokens
2. Accept the conditions at https://huggingface.co/pyannote/speaker-diarization-community-1
3. Put the token in .env at the repo root: SCRIBE_HF_TOKEN=hf_...

Without a token scribe still transcribes, it just cannot tell speakers apart.

Usage

CLI

.venv\Scripts\python -m scribe.cli transcribe path\to\audio.mp3 --diarize -f srt -o out\

Web UI and API

Run the server (python -m scribe.server) and open http://localhost:8323 for the drag-and-drop UI, or drive it programmatically: upload a file, poll the job, then fetch the result in any export format.

Configuration

Every setting is an environment variable, or a line in .env at the repo root:

Variable	Default	Meaning
SCRIBE_MODEL	large-v3	Whisper model size (tiny ... large-v3)
SCRIBE_DEVICE	cuda	cuda or cpu
SCRIBE_COMPUTE	int8_float16	CTranslate2 compute type
SCRIBE_HF_TOKEN	unset	Hugging Face token for diarization (falls back to HF_TOKEN, then the standard Hugging Face token cache)
SCRIBE_DIARIZATION_MODEL	pyannote/speaker-diarization-community-1	Diarization pipeline to load
SCRIBE_HOST	127.0.0.1	Server bind address; set 0.0.0.0 to expose on the network
SCRIBE_PORT	8323	Server port
SCRIBE_DATA_DIR	./data	Uploads, results, job database
SCRIBE_MODELS_DIR	./models	Model cache (HF_HOME points here)

Contributing and security

See CONTRIBUTING.md for setup and checks. Report vulnerabilities privately per SECURITY.md.

License

MIT, see LICENSE.