As darling of data science Andrew Ng said, "90% of machine learning is feature engineering". So you should probably run the STFT spectrograms through machine le

As darling of data science Andrew Ng said, "90% of machine learning is feature engineering". So you should probably run the STFT spectrograms through machine learning models rather than the raw audio signal. And anyway raw audio is over 41k samples per second so quite low information density compared to some simple and quite common transformations like STFT anyway.