Hi,
In my project I have 3 labels of datasets, in my model Classifications and performance is good but when I’m running Edge impulse model on “rpi4 keyword spotting” is cool but, when I’m giving keyword and some audio parallelly keyword is not spotting and getting false spotting. Any better suggestion to stop false keyword spotting? thanks!
project ID :718166
Hello @Bharath_amitekh,
Is your dataset representative of the real conditions? Like background noise mixed with your keyword? Other/Unknown class that can catch the false postives?
Is the microphone used to record the data the same as the one used to recognize your words (or at least with the same parameters (gain/frequency/etc…).
One easy thing you can do is to try to record some samples from your raspberry and then run the Data Explorer with your trained impulse.
If you see that your newly collected data (from the RPi) are far off the clusters, it would indicate that the audio samples from your training dataset and your raspberry input buffer are different.
Here is a quick script that can help you collect more data using your model from your RPi:
Best,
Louis
Hello @louis ,
Thank you for your response.
Yes, the dataset is in good condition and appears to be representative. However, I did notice that some background noise has mixed with the keyword samples.
Also, I’m using the same microphone for both recording the dataset and recognizing the keywords during inference.
Best regards,
Bharath.