Hi JTedot,
For coherence raison, default tflite Yamnet model embeds also the preprocessing stage of the raw audio samples to build the features to feed a "classical" deep network. This part is exported as a set of "basic" TFLite operators (Mul, Cell, Sub, Pad,...), this is in theory interesting for the end-user but to deploy it efficiently on the IoT target, this is a little bit challenging. It is recommended to use the specialized and more efficient libraries to compute the features and after to deploy/generate the next part. I have not checked, but I think that the original scripts to generate the whole model can be updated to generate only the second part and to extract the used parameters for the first part. Or your experimentation to fix the dynamic input size is correct (and recommended to deploy a model, dynamic input size is not supported) but don't forget the first batch dimension: (1, nb_samples) instead (nb_samples, ).
Concerning the "1-D" issue, X-CUBE-AI supports the 1-D tensor. The issue here is that the model has an input shape with only one dimension. X-CUBE-AI expects a minimum of 2 dimensions. First is considered as the batch dimension, for a 1D-tensor, the tools expects the input shape; (1, nb_features).
br,
Jean-Michel
