Self-trained models seem to work. Check out this article: https://www.kdnuggets.com/2020/02/audio-data-analysis-deep-learning-python-part-1.html
At the end, you could convert your model via
import tensorflow as tf
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
open("my_model.tflite", "wb").write(tflite_model)
and pass that to the X Cube AI. It passes the analysis correctly. I'm thinking about dropping YAMNet on embedded systems, since too many functions happen abstractly under the hood and aren't really accessible. If you build your own model, you are in full control, but you also have to extract features yourself. You can't pass a network an audio file, you have to pass it a highly filtered spectrogram (in essence audio classification is image classification).
