cwpearson / scene-scribe

Scene Scribe

Getting Started

• Get an access token at https://huggingface.co/settings/tokens.

• install ffmpeg

pip install -r requirements.txt

HF_TOKEN=python -m transcribe path/to/video.mp4

• Diarizing consumes about 2.7 GB of GPU memory.

Todo

does condense_segments need to happen where it is? It rejoins segments split by scene.

Basic Idea

Use yt-dlp to retrieve video from CABQ website

Use ffmpeg to extract audio and convert to .wav.

ffmpeg -i input_video.mp4 -vn -acodec pcm_s16le -ar 44100 -ac 2 output_audio.wav

Use whisper for transcript

Use pyannote.audio to diariaze

Use ffmpeg to extract a screenshot from the video for each speaker turn

ffmpeg -ss HH:MM:SS.ms -i input_video.mp4 -frames:v 1 output_screenshot.jpg