I’m using the Nicla Voice board and working on a project called Audio2 (Edge Impulse Project ID: Audio2). I developed my model using the Edge Impulse Studio, and during Live Classification, the model performs very well — predictions are accurate and consistent.
However, after deploying the model to the board and running it with edge-impulse-run-impulse --debug, I consistently get wrong classification results. The predictions no longer match the behavior I saw during Studio testing.
Has anyone else experienced this mismatch between Studio and on-device inference? Could this be related to differences in audio processing, sampling rate, or normalization between the two environments?
Any insights or suggestions are appreciated!
Thanks in advance.