Exploring Multi-Descriptor Disentangled Representations of Acoustic Instrument Notes

Comprehensive representation is key for improving controllability in generative neural networks. We present an approach for learning disentangled latent representations of individual instrumental notes, leveraging a Variational Autoencoderbased architecture designed to operate on spectrograms and explicitly capture four musical descriptors: timbre, pitch, dynamics, and duration. To achieve a structured and interpretable latent space, we exploit a combination of Gaussian Mixture priors, adversarial training, and auxiliary supervised clustering, promoting both compactness and semantic coherence in the learned representations yet preserving the ability to accurately reconstruct the original spectrograms. Experimental results and latent space explorations on the TinySol dataset show the effectiveness of the proposed approach, outperforming baseline models and existing methods in key metrics of reconstruction quality and classification accuracy.