Computational Creativity by Structural Analogy between Acoustic Bayesian Models

We present a method for generative modeling of audio content that performs mappings between minimum entropy hidden Markov models learnt from audio data. By training with a minimum entropy prior, compact, low-complexity models of the latent structure in audio source samples are obtained. Synthesis of new audio content is achieved by mapping the state sequence of a nominated structure model onto a nominated content model. The mapping is chosen such that the cross-entropies between the state variables of the nominated models are minimised. This creates an analogy between the models’ structures, even when the specific content of the models varies significantly. The re-mapped content state sequences are inverted to yield spectral vectors that consist of the higher-order pattern information of the structure model and the low-order spectral features of the content model. To illustrate the methods, we present examples of mapping audio structure and content between drum beat samples in different styles.