I am following this tutorial for audio classification. My results are correlated with the author's results and I am getting $99\%$ accuracy for the train set and $91.8\%$ accuracy for the test. I have omitted the averaging code line
mfccsscaled = np.mean(mfccs.T,axis=0) and got similar results.
I want to deal with the over-fitting gap of $\sim7\%$. I have tried a higher drop out rate, a layer l2 regularization for wither the cnn layers, the dense layer, or both with numerous values, and I have tried combinations of normalization and dropout. None of these methods worked.
The reason I am attending this is that a number of colleagues have stated that this is an un-acceptable overfit gap. Is that really the case, or are these normal results that cannot be improved?
I read a lot of information online on over-fitting. I am limited to Urbansound8k dataset which means that more data as a method of dealing with over-fit is not an option. Also, I know that today's audio classification methods use CRNN which takes longer to train. I am satisfied with the tradeoff of lower results with CNN over higher results with CRNN but the overfitting problem is unrelated + I understand that a smaller network has less tendency to over-fit. Is there any other method I am missing for reducing this gap?