Transcripts

How it works

Each person’s audio is transcribed on its own, with the language detected per speaker, so speakers are never confused. The lines are then merged by time into Transcript.txt, .docx, .srt and .vtt with names. They appear a little after the recording’s audio and video files.

Model and speed

The Docker image ships Whisper base (multilingual, 148 MB) and transcription is on by default there. For better accuracy, build with --build-arg WHISPER_MODEL=small and its WHISPER_MODEL_SHA256 (466 MB, about three times slower). It runs on CPU; WHISPER_THREADS (default 4) sets how many cores one transcript uses.

# deploy/compose/.env: turn transcripts off
WHISPER_MODEL=

docker compose up -d