Transcribe audio to text (using whisper). Runs locally.
Files: transcribe.py, youtube_util.py. Possibly select start and end time.
Optional UI (served by flask) using python -m src.app to:
- Prepare transcript from a YouTube link or a file.
- Text search the database (MongoDB) of transcripts. Note: stop words (like 'how', 'is', 'why') are not indexed.
- Edit saved transcripts.
Takes a few seconds for to transcribe a few minutes of audio, for example this song took less than 5 seconds on my pc using cuda:
File: transcribe.py. Output example:
File: speaker_diarization.py, separating the speakers is done using pyannote (hugging face token required). Output example:
SPEAKER_01: No introduction needed.
SPEAKER_03: Welcome. I just agreed to this last minute, as you know.
...
Install and read the instructions for torch, pyannote (for separating speakers. hugging face token required), and install the other requirements.
Set device ID's in settings.py in order to record audio (helper functions are in audio_util.py).
For real-time transcription - there are probably better ways, but it works surprisingly well for short and fast transcriptions.
The model may hallucinate a bit, and be non-deterministic.