Keyword Spotting tips to increase unknown accuracy


I am a beginner in keyword spotting ML.
I have followed the tutorial and successfully integrated the C++ library generated by EI to my STM32 Sensortile board. Now the my task is to understand how to improve the prediction accuracy and reduce false prediction.

Looking at the graph below, even the accuracy said 93%, there is >10% that unknown voice can be predicted as “One” or “Zero”, which will make the demo kind of look really bad.

Any tips on reducing false prediction on Unknown data? Or the only way is to look at the visualization graph and delete manually one by one?

Thank you

Hi @hkartadi,

As you mentioned you have lots of false predictions in your Unknown class. I think your “one” and “zero” classes don’t have enough data. We advice to capture at least 10 minutes of audio for each class to have good performances. You can also start by having 5 minutes for each and compare the performances.