Efficient variational inference for the dynamic harmonic model

In this paper, we develop a class of probability models that are potentially useful for various music applications such as polyphonic transcription, source separation, restoration or denoising. This class unifies and extends several models such as sinusoidal and harmonic models, additive synthesis model, Gabor regression and probabilistic phase vocoder. We overcome computational intractability issues by introducing structured variational (mean-field) approximations that lead to efficient local message passing algorithms