I followed the turtorial “https://docs.edgeimpulse.com/docs/custom-blocks” to realize ei_keyword_spotting application.
I finished adding custom block on the studio, but the training accuracy is about 0.29, while using mfcc block on edge impulse ,the training accuracy is about 90%.
Meanwhile, when I use mfcc block, the number of input features is 650, when I use custom block, the number of input features is 16000, how to adjust it?
Have you implemented the custom block given in example in the documentation? The example implements a smoothing algorithm which is typically useful for motion detection, not for audio.
Which audio algorithm do you need to implement?
Hi @aurel , thanks for your reply!
I used mfcc on github to inplement the custom block, which is as shown in the figure
which caused the above problems happened
650 features looks correct for a 1 second window and default MFCC parameters, I just gave it a try with the custom block from github and it outputs 650 features too.
Do you see the following parameters with the custom block?
@sherry Looks like your custom block is based on the raw data block (given the 16000 features you mention), so if you could share how you’re starting your server that’d be helpful.