Novel Generative Classifier for Acoustic Events

In this paper, we describe a novel generative classifier for audio events based on the projected belief network (PBN). The PBN is a layered generative network formed from a feed-forward network, so it can be simultaneously trained as a generative model for a given class, and as a discriminative classifier against "all other classes". A PBN can also be shortened to any number of layers, allowing the output features of the shortened network to be modeled using an arbitrary probability density function (PDF) estimator. We exploit these properties of PBN to model the features from the output of an early convolutional layer, where a time dimension is still present, using a hidden Markov model (HMM). The special generative/discriminative training of the PBN produces generative features that are also high in discriminative information, forming a generative classifier combining (a) a discriminative deep network and (b) a generative neural network, and (c) a HMM classifier rooted in classical Bayesian approaches. The approach is demonstrated in the task of acoustic event classification.