Hi there. I am experiencing problems identifying simple drops in a cloned project that works essentially well without them added !
With drops recordings added and model retrained untouched, i get a perfect confusion matrix and classes repartition, as this can be seen in the cloned project :
But, when running a live classification on a live 5 seconds recording carried-out from my desktop with absolutely not water dropping around, my recording is identified as being a drop 100% of the sample.
Could one have a look on this project and help me understand what i am missing here?
Many thanks for your help