Local Hebrew Transcription Script

This project uses faster-whisper and the Ivrit.ai model to transcribe audio/video files.

• uv (An extremely fast Python package installer and resolver)

1. Make sure uv is installed.

2. Clone this repository (if not already done).

3. The environment and dependencies will be automatically handled by uv when you run the script.

4. Create a .env file in the project root to store your configuration:

   HF_TOKEN=your_hugging_face_token

   Note: The HF_TOKEN is required for speaker diarization (the process of separating audio into different speakers). It is used only to verify that you have accepted the user agreement for the pyannote/speaker-diarization-3.1 model on Hugging Face. No sensitive data is accessed or shared. You must accept the agreement here: https://huggingface.co/pyannote/speaker-diarization-3.1

To run the transcription script, use uv run:

uv run transcribe.py path/to/media_file

To transcribe all media files in a directory:

uv run transcribe.py path/to/folder

Supported extensions: .mkv, .mp4, .m4a, .mov, .amr, .wav, .mp3, .flac, .opus.

Simple example:

uv run transcribe.py interview.mp3

Example with diarization (different speakers identification):

uv run transcribe.py --diarize interview.mp3

Note: Transcription outputs are saved as .txt files in the same directory as the source media.

⚠️ Note: This script was primarily "vibe-coded" and is intended mainly for personal use. It has not been rigorously tested for production environments. Use at your own risk.

This project relies heavily on the incredible work of Ivrit.ai.

Ivrit.ai is a non-profit organization dedicated to making Hebrew AI accessible to everyone. The default model used in this script (ivrit-ai/whisper-large-v3-turbo-ct2) is one of their state-of-the-art models. We deeply appreciate their contribution to the open-source community and their mission to advance Hebrew language technology.

This project is licensed under the MIT License - see the LICENSE file for details.