Experiments with speech compression/processing. The modern approach to this (and pretty much everything else) is to transform the input data into a high-dimensional space, throw it into a neural net and cross your fingers. (the 'Whisper' STT runs in 512-dimensional space!). It works, but it's hardly elegant or compute-efficient. There's a certain art in doing things the 'old way', even if a million years of evolution decided neural-nets were the one true path.
replies(1):