Evaluation of a Feature Compensation Approach Using High-Order Vector Taylor Series Approximation of an Explicit Distortion Modelon Aurora2, Aurora3, and Aurora4 Tasks

Abstract

In our previous work, a new feature compensation approach to robust speech recognition was proposed by using high-order vector Taylor series (HOVTS) approximation of an explicit model of distortions caused by additive noises, and evaluation results were reported on Aurora2 database. This paper extends the above approach to deal with both additive noises and convolutional distortions, and reports evaluation results on Aurora2, Aurora3, and Aurora4 tasks.