In creating my text data for a whisper fine-tuning, can I use “<|nospeech|>” directly as the text? (This is for my wife, who’s breathing with a ventilator – I want to ensure the ventilator sound is properly trained as being NOT speech.)
Also, while I plan on having spans of noise mapped to the nospeech token, is it possible to put it amidst a single sentence?
For instance, if the audio has sentences with gaps in it, like for normal spoken delays: “Hmm… [brief pause]… Yeah I believe so.” Is <|nospeech|> suitable to place in the pause?
[Edit] This is the script I’ve been using to fine-tune a whisper/hf model with my own recordings/transcriptions of my wife: