Skip to content

Initial commit to support speech to text via openai.whisper.

Paul S Hahn requested to merge PaulSHahn/recoll:speech_to_text into master

Add optional speech to text transcription support for the audio input handler rclaudio.py.

If option speechtotext = whisper appears in the recoll config file, transcribe's audio output using page breaks to represent seconds. Only do so if lyrics are not already tagged in the metadata.

Opening a snippet in recoll results in the player going to the second where the audio was transcribed in vlc (approximately).

Configuration option sttmodel is used to specify the language specific compute model used to train the transcriber against.

Requires Python handlers openapi-whisper and filelock. Filelock syncs access to the parallelized compute resource to one process at a time, which keeps concurrent indexers from running out of compute resources.

Merge request reports