Improving Interpretability in Generative Multitimbral DDSP Frameworks via Semantically-Disentangled Musical Attributes

We propose a generative multitimbral DDSP framework focusing on acoustic instruments modeling. Traditional DDSP architectures rely on frame-wise latent encoding, limiting semantic interpretability; the acoustic realm is further challenged by the inherently interdependency of musical descriptors. Our approach introduces disentangled global embeddings for timbre and dynamics, achieved via a Triple-VAE architecture, and a DDSP model to learn temporal dependencies from such reduced representation along with normalized envelope curves. Experimental results demonstrate optimal reconstruction accuracy while enabling flexible and realistic generation capabilities, with the disentangled representation promoting interpretable control over musical descriptors.