Key Model

[tomxi]

I would like to build a key model.

The model will take audio as input, run it through CREMA's chord model as a feature extractor, and take the structured chord outputs as features to train a CREMA style network.

The network will predict a structured output for keys:(tonic, pitch_profile) where pitch_profile is a 12-D vec (1 field per pitch class); it'll also use the above encoding of a key to predict a key tag. All in the style of CREMA chord model.