Comment by tgv

By the looks of it, those functions extract features (like frequency peaks). You do that once for a sound. The output could function as input for an NN, in which case it would be a tokenizer for sound.