Fusion of parametric and non-parametric approaches to noise-robust ASR

In this paper we present a principled method for the fusion of independent estimates of the state likelihood in a Dynamic Bayesian Network (DBN) by means of the Virtual Evidence option for improving speech recognition in the aurora-2 task. A first estimate is derived from a conventional parametric Gaussian Mixture Model; a second estimate is obtained from a non-parametric Sparse Classification (SC) system. During training the parameters pertaining to the input streams can be optimized independently, but also jointly, provided that all streams represent true probability functions. During decoding the weights of the streams can be varied much more freely. It appeared that the state likelihoods in the GMM and SC streams are very different, and that this makes it necessary to apply different weights to the streams in decoding. When using optimal weights, the dual-input system can outperform the individual GMM or the SC systems for all SNR levels in test sets A and B in the aurora-2 task.