Bayesian Spectral Matching: Turning Young MC into MC Hammer via MCMC Sampling
Matthew D. Hoffman, Perry R. Cook, David M. Blei · 2009
In this paper, we introduce an audio mosaicing technique based on performing posterior inference on a probabilistic generative model. Whereas previous approaches to concatenative synthesis and audio mosaicing have mostly tried to match higher-level descriptors of audio or individual STFT frames, we try to directly match the magnitude spectrogram of a target sound by combining and overlapping a set of short samples at different times and amplitudes. Our use of the graphical modeling formalism allows us to use a standard Markov Chain Monte Carlo (MCMC) posterior inference algorithm to find a set of time shifts and amplitudes for each sample that results in a layered composite sound whose spectrogram approximately matches the target spectrogram. 1.