Investigate pre-processing options to detect long silences in recordings to be transcribed via Whisper

[andrewjbtw]

Whisper has a tendency to generate hallucinations over periods of silence in recordings. This ticket is to investigate possible options for pre-processing audio to see if that can lead to an improvement in Whisper output.

This recording that's been accessioned in QA could be a good sample since it starts without about 17 minutes of silence and has a few other periods of silence.

This case is distinct from:

• recordings that have no audio track whatsoever
• recordings that have an audio track but are essentially silent (below a certain audible threshhold for the whole recording)

Those two other cases are more tractable and higher priority (see #1436).

[jmartin-sul]

Warning from @dnoneill at storytime: when pre-processing input into chunks and omitting silent chunks, cutting out silent segments that are shorter than a certain threshold (5 seconds?) can make Whisper results worse for subsequent non-silent content. So we only want to cut silences of more than a certain length when segmenting videos we send to whisper. Will likely require some tinkering/tuning for anything more than detection of silence in general (i.e. case 2 in @andrewjbtw's description), so maybe not enough time in this work cycle to complete and refine this sort of pre-processing step.