I made my own model to regocnise 20 commands. It gets words right pretty well but seems like there is no threshold for predoction. Seems like if it does not understand the Word it chooses one at random from vacabulary list. Is It possible to somehow get out the accuracy value per prediction and set prediction threshold using mic_vad_streaming.py for example?
Use the metadata function and you’ll get a confidence score. Not exactly accurate, but it should give you some indication.