Transcription Friend
Turn interviews, voice memos, and meetings into clean, timestamped text with GPU-accelerated Whisper. Built for batches of m4a segments, with Danish and English support and language-specific corrections.
Transcription runs on your own machine, so your audio stays private. This page is a preview of what it produces — grab the full tool on GitHub to run it yourself.
Sample
[00:00:00] Velkommen til samtalen. Kan du starte med at fortælle lidt om dig selv? [00:00:06] Ja, selvfølgelig. Jeg har arbejdet med digitalisering i den offentlige sektor i over ti år. [00:00:13] Og hvad fik dig til at interessere dig for det område? [00:00:18] Det startede med et lille projekt, hvor vi skulle automatisere en manuel arbejdsgang.
Illustrative output. The tool also writes a corrections report showing any language-specific fixes it applied.
What it does
- Engine
- faster-whisper (CTranslate2)
- Quality presets
- Fast · Balanced · Best (tiny → medium)
- Languages
- Danish, English, + auto-detect
- Output
- Plain text + timestamped, optional corrections
- Acceleration
- NVIDIA GPU (CUDA), CPU fallback
- Formats
- m4a, mp3, wav, mp4, flac, ogg…
Run it locally
Clone the repo and start the local drag-and-drop interface. Your recordings are processed on your machine and never uploaded.
git clone https://github.com/Walliiee/transcription_friend cd transcription_friend pip install -r requirements.txt # local drag-and-drop UI at http://127.0.0.1:8000 python scripts/web_api.pyTranscription Friend on GitHub
Frequently Asked Questions
A GPU-accelerated audio transcription workspace built on faster-whisper. It turns m4a (and mp3, wav, mp4, and more) recordings into clean text — both plain and timestamped — with Danish and English support and language-specific post-processing corrections.
No — transcription needs a Whisper model and ideally an NVIDIA GPU, which a static website cannot provide. This page shows what the tool produces; to transcribe your own audio you run it locally on your machine, where your recordings never leave your computer.
A GPU makes it fast, but it falls back to CPU automatically (just slower). For the newest GPUs (e.g. RTX 50-series, sm_120) install a CUDA 12.8+ build of PyTorch. See the repo SETUP guide for details.