Overview

Is it possible to use directional filters instead of square filters for key/tempo detection?

Motivation

Architecture

Results

Further Thoughts

It seems it is indeed possible and for small networks as long as the directionality of the filters matches what we're looking for it is going to work. However, square filters work in both cases and are easier to think about and implement. However, the good results of the square filters might be due to confounding the tempo/key with genre information. Research wise this is a problem, but in real applications I imagine it doesn't matter as much.

Links