Hi,
In the Edge Impulse Studio, there is a possibility to change the preemphasis coefficient, which has a default value of 0.96875. I understood that this has an impact on the spectrogram (generated features).
In some cases, I can achieve much better accuracy of the ML model in the Edge Impulse Studio when I change this value significantly (e.g., to 0.1). However, in this case, I observe a very different behavior after deploying the model into the TinyML board (which does not correspond to the training/testing accuracy at all). It seems to me that this might be caused by the physical chip having this coefficient fixed and not being able to be changed, even if I change it in the Edge Impulse Studio. Can you please confirm or deny that changing the preemphasis coefficient in the Edge Impulse Studio will change this value even after deployment into the NDP101 on the TinyML board?
Hi,
Changing the preemphasis coefficient in Edge Impulse Studio does not change the value on the NDP101 (TinyML) after deployment.
The pre-emphasis coefficient you set in Edge Impulse Studio defines a pre-processing step applied to the audio data during training and simulation. This filtering is not directly embedded into the model itself.
Thank you.