xbasset
commented
4 months ago The point is to control the "volume" of vocab parameters sent to Whisper prior to the call.
Few options to explore:
- have a big list of vocab per user in database and filter it depending on the context to populate a fixed sized list to whisper parameter
- deduplicate / filter / rank vocabulary per user to keep it clean and organized per context.
- hope whisper API would accept ~infinite vocab list as input...
Mojodex can learn only 245 tokens as "vocabulary" for transcription. We should find a way to lift the restriction.