Two-domain feature compensation for robust speech recognition

Abstract

In this paper, we develop a two-domain feature compensation approach to the log-filterbank and log-energy features for reducing the effects of noise. The environment model is approximated by statistical linear approximation (SLA) function. The cepstral and log-energy feature vectors of the clean speech are trained by using the Self-Organizing Map (SOM) neural network with the assumption that the speech can be well represented as multivariate diagonal Gaussian mixtures model (GMM). With the effective training of clean speech and environment model approximation, noise statistics is well estimated using batch-EM algorithm in a maximum likelihood (ML) sense. Experiments in the large vocabulary speaker-independent continuous speech recognition demonstrate that this approach exhibits a noticeable performance. © Springer-Verlag Berlin Heidelberg 2005.