Training strategies for deep latent models and applications to speech presence probability estimation

In this study we address models with latent variable in the context of neural networks. We analyze a neural network architecture, mixture of deep experts (MoDE), that models latent variables using the mixture of expert paradigm. Learning the parameters of latent variable models is usually done by the expectation-maximization (EM) algorithm. However, it is well known that back-propagation gradient-based algorithms are the preferred strategy for training neural networks. We show that in the case of neural networks with latent variables, the back-propagation algorithm is actually a recursive variant of the EM that is more suitable for training neural networks. To demonstrate the viability of the proposed MoDE network it is applied to the task of speech presence probability estimation, widely applicable to many speech processing problem, e.g. speaker diarization and separation, speech enhancement and noise reduction. Experimental results show the benefits of the proposed architecture over standard fully-connected networks with the same number of parameters.