Question/Issue: I am in the midst of developing a keyword spotting device. To this end, I’m following this guide: On your Espressif ESP-EYE (ESP32) development board - Edge Impulse Documentation
I have cloned the repository and deployed my model, which is a mobilnet v2.
In order to test it I am copying raw data from the live classification tab. In my case it is an array of 16000 samples from a .wav file with the keywords.
The problem I am having is that, no matter the file I pick, the prediction is always bad, I have tested the optimised int8 model and the unoptimised float model.
When I test the int8 model the prediction is always the “unknown” label with a 99.6% score.
In contrast, when I test the float model, the prediction changes a bit, but it is always the “unknown” label but the score varies from 55% to 99%.
Am I doing something wrong with this example? Is there a problem with it?
Thank you for any help.
Project ID: 265677
Context/Use case: keyword spotting on the end device