Offline Transcriber

Local Python app that turns MP3 (and other audio) into .txt and .srt files using faster-whisper. After a one-time model download, it runs fully offline. No cloud APIs, no accounts, no telemetry.

Command line and a simple desktop window are both included.

Requirements

• Python 3.10+ (3.11–3.13 recommended)
• Disk space for a model (small is a few hundred MB; large-v3 is a few GB)
• Optional: NVIDIA GPU + CUDA for faster / larger models

Works on Windows, macOS, and Linux.

One-time setup (internet)

git clone https://github.com/akgordon/Transcriber.git
cd Transcriber
python -m venv .venv

Activate the virtual environment:

• Windows (PowerShell): .\.venv\Scripts\Activate.ps1
• Windows (cmd): .venv\Scripts\activate.bat
• macOS / Linux: source .venv/bin/activate

If PowerShell blocks the activate script:

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
.\.venv\Scripts\Activate.ps1

Then:

pip install -r requirements.txt
python -m transcriber download small

That stores weights in models/small/. Repeat download with another size if you want a second model.

After this, you can disconnect from the network. Transcription never contacts Hugging Face or any other service.

Use it

Desktop window:

python -m transcriber --gui

Command line:

python -m transcriber path/to/audio.mp3
python -m transcriber path/to/audio.mp3 --model small --language en
python -m transcriber path/to/folder --model small
python -m transcriber download small

Outputs are written next to the source file:

• audio.txt — plain transcript
• audio.srt — timestamped subtitles

Supported files: .mp3, .wav, .m4a, .flac, .ogg, .opus, .webm. Folder mode is non-recursive.

Options

Flag	Meaning
--model	tiny, base, small (default), medium, large-v3, turbo
--language	auto (default) or a code such as en
--device	auto (default), cpu, or cuda
--gui	Open the desktop window

On CPU, start with small. Use large-v3 or turbo if you have an NVIDIA GPU.

Offline rules

• Models load from models/<name>/ with local_files_only=True.
• If a model is missing, the app tells you to run python -m transcriber download <name> instead of downloading during transcription.
• Silero VAD (silence skipping) is bundled with faster-whisper; it does not need a network.

Contributing

See CONTRIBUTING.md. Please report security issues as described in SECURITY.md.

License

This project is licensed under the MIT License. Copyright (c) 2026 Alan Gordon.

Third-party components:

Piece	License
faster-whisper	MIT
CTranslate2	MIT
OpenAI Whisper weights	MIT
Silero VAD	MIT
FFmpeg / PyAV (MP3 decode, installed via pip)	Typically LGPL

This project does not call the paid OpenAI transcription API. Model weights are not shipped in the repository; each user downloads them locally.