5 comments

[ 3.8 ms ] story [ 21.0 ms ] thread
Just noting that this is long and incredibly basic. It doesn't cover anything about spectrograms for example despite the subject title.
There needs to be experiments with representations such as chirplets since those are actually physically relevant.
(comment deleted)
FWIW, if you want to dip your toes into audio and song classification, the Spotify API offers feature extraction.

Also, there are some very cool ML topics in the frequency domain. For example - in the past years there's been a small revolution in the guitar / bass / etc. industry where simulators offer quite realistic sounding amps/speakers sounds.

Back in the day, these simulators were just carefully crafted filters that would emulate the response of said amps, speakers, room/ambience, etc. but with the advent of more powerful DSP, it is now possible to sort of "extract" these features on the fly - some of these products do indeed use ML/DL algorithms in their core. In the end, the goal is to profile the environment, and create some complex filter / function that matches the response. Which is one thing Neural Networks can do quite well (approximating functions).