I know this isn't new, but I started this from scratch organized about 60 genre's manually. It classifies music based on patterns from chunks of the spectrogram up to 16khz. It's pretty accurate on electronic and rock subgenres and is still being tweaked. It's serverless on the classifier microservice so it may be slow if noone else is using it at the runpod scales to 0 if noone is using it for cost savings.