Device inference underperforming site inference

I posted about this here but figured I should create a new topic. Based on the info from the linked topic, I’m attempting to run continuous inference on a nrf52840. Unfortunately the values I’m getting back are significantly less than what I get with the online classifier. Printing out the results, it sort of feels like the event are occurring on slice boundaries although my understanding was that continuous inference should help mitigate that phenomenon. Once in a while I get high values (e.g. 0.99) for my keyword. Other times, nearly zero. In contrast, uploading a clip to the online classifier gives me reliably > 0.8 in detecting keyword events.

Hi @jefffhaynes, can you remove the moving average filter as described here: ?

Definitely an improvement! Still not to the level I’m seeing online but probably good enough…I’ll play around a bit with averaging and see if I can boost it a bit.