Transcribe audio on your own machine

Drag a recording onto a web page, and Whisper transcribes it on the workstation it runs on — with speaker labels, live progress and exports. Anyone on your network can use it; nothing leaves the network.

Download for Windows View on GitHub

Windows installer with a tray app, or run it from source on Windows, Linux or macOS with uv.

A finished transcription card: two named speakers, Daniel and Samantha, with their talk time, a timestamped transcript labelled by speaker, and buttons to copy, save as txt, srt, vtt or json, retry and relabel speakers. — A finished job: speakers named on the card, transcript labelled line by line, five export formats.

What it does

Drop, set, transcribe

Stage audio or video files, choose the model, language and quality, and press Transcribe. Uploads show their progress, and the transcript streams in live as each segment finishes.

Who said what

Speaker identification runs locally with sherpa-onnx — no PyTorch, no account. Set the speaker count, then name speakers or merge a stray one on the card, or relabel without transcribing again.

One file per person

Zoom can record a separate audio file for each participant. Drop them together as one call and every line is attributed exactly, named after the person, with no guessing.

Exports that fit

Copy the text, or save plain text, timestamped text, SRT, WebVTT with speaker voices, or JSON with segment and word timings.

Private by design

An access token guards the page. A tamper-evident audit trail, behind a token of its own, records what happened — hashes, never transcript or prompt text.

Made for the GPU you have

With --preload it proves the card can really encode before it calls itself ready, and when CUDA is not usable it says what to fix. CPU works too, more slowly.

Get started

Windows: the tray app

Download TranscriptionServer-…-setup.exe from the latest release.
Run it. It is not code-signed yet, so SmartScreen will ask you to confirm; SHA256SUMS.txt on the release lets you check the download.
Start it from the tray. It picks the port, holds the access token, and opens the page for you.

Any OS: from source

The server is one Python script that declares its own dependencies, so uv is the whole install:

git clone https://github.com/yorch/whisper-transcribe-server
cd whisper-transcribe-server
uv run transcribe_server.py --preload

Open the link it prints. Add --device cpu on a machine without an NVIDIA GPU. ffmpeg is needed for anything that is not plain WAV.