A Generative Model of Impulsive Sound Production and Propagation
The overall objective is to design a learning system that takes a training set of acoustic signals and produces a classifier that can identify the category of the acoustic signal, out of a small number of categories. We have already found that it is important to consider the factors of variation that can influence the signal, and the proposed project aims at exploiting physical modeling knowledge to structure a generative model of the acoustic production of the observed signals. We propose is to develop a probabilistic generative model that will detail the effect of sound propagation through the atmosphere, including in particular the effect of interaction between the source signal and the ground. We plan on leveraging knowledge of the physics of atmospheric sound propagation as well as our knowledge of the structure of the target sound sources to improve classification accuracy. A very important aspect of this project is the development of accurate and computationally efficient inference schemes. Since it is likely that the resulting model will not be amenable to exact inference, we will require effective approximate inference schemes that permit the joint estimation of various factors contributing to the observed data, including the source class, the ground type, the distance from the source to the microphone array and the orientation of the microphone array. The approach we will adopt will be to perform exact inference over those aspects of the model where this is possible, for example, the source generation model will consist of a mixture of Gaussians that is amenable to exact inference. However, over more inferentially intractable aspects of the model, such as the propagation transfer function, we intend to explore both sampling (via MCMC) schemes and deterministic variational approximations.
