Hello,
I have trained a keyword spotting model using Edge Impulse with two labels: “temperature” and “noise”. Both classes have an equal number of samples. However, after uploading the generated library to the Arduino IDE, the model gives predictions even when no one is speaking into the microphone.
12:11:04.399 → Predictions (DSP: 133 ms., Classification: 6 ms., Anomaly: 0 ms.):
12:11:04.399 → Temperature: 0.66797
12:11:04.399 → Other: 0.33203
12:11:04.399 → Starting inferencing in 2 seconds…
12:11:06.408 → Recording…
12:11:07.340 → Recording done
12:11:07.469 → Predictions (DSP: 133 ms., Classification: 6 ms., Anomaly: 0 ms.):
12:11:07.469 → Temperature: 0.00781
12:11:07.469 → Other: 0.99219
12:11:07.469 → Starting inferencing in 2 seconds…
12:11:09.480 → Recording…
Hi @Andalib
You either have too few keywords in the dataset or you are having a hardware fault with the microphone. Can you record audio to the project via the Arduino?
See the dataset section for details on how to collect the correct amount:
Best
Eoin