This project provides a Python-based pipeline for extracting audio from video files, performing speaker diarization, and transcribing audio using OpenAI's Whisper. It identifies different speakers in the audio and aligns transcriptions with these speakers.
- Audio extraction from video files
- Audio trimming to a specified duration
- Speaker diarization to identify different speakers in the audio
- Audio transcription using OpenAI's Whisper
- Matching transcriptions with speaker diarization data
- Python 3.x
- Libraries:
pydub,yt_dlp,pyannote.audio,whisper - ffmpeg (for audio extraction from video)
- Ideally CUDA GPU with 10GB for whisper large or 6-8gb for whisper medium - will run on CPU, just slower.
- See docker file
Ideally docker compose a dev environment via the compose yaml / dockerfile
Usage Configuration is made in config.ini - there you can set a path to your input directory (where audio or video files are stored). Make changes (other than the model) to the transcription settings within transcript.py. Run the main.py script.
audio_processing.py: Handles audio extraction and trimming, noise reduction, normalization, produces a mp4 with embedded subtitles. speaker_diarization.py: Performs speaker diarization using pyannote.audio. transcription.py: Transcribes audio using Whisper. main.py: Integrates all modules and runs the pipeline - produces various output formats.
License [Your chosen license]
Acknowledgments OpenAI for providing the Whisper model. Contributors and maintainers of pyannote.audio, pydub, and yt_dlp.