Is it possible to use "prompt" or "hotwords" to steer decoding similar to Whisper?
^title
It should be possible to do with a corpus at least: https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/asr/asr_customization/ngpulm_language_modeling_and_customization.html#ngpulm-ngram-modeling
Training the n-gram model is really fast.
But any time I add it and try to transcribe on a sound file that's more than 40 seconds long (still less than a minute) it will drop a bunch of sentences.
And then they have word boosting which doesn't seem to work on an AER model like canary
This function should be in by default. For the most useful use cases, business meetings or so, concrete vocabulary is almost always needed.
I would love to see v3 with --prompt included, because now I have to stick with Whisper (small) just for that reason when Canary makes better results with x2 speed.
I think the issue for me was that I concatenated data to get longer annotated data but the volume was vastly different for the sound files.