Robust speech recognition under noisy environment using speech rate training system

Abstract

This paper proposes a speech rate training system approach to characterizing environments to improve the performance of automatic speech recognition system under noisy environment conditions and speaking rate differences. The speech rate training system consists of two phases, the offline and the online. In the offline phase, a speech rate training system is formed by a collection of super vectors. Each super vector consists of the set of means of all the Gaussian mixture components of a set of HMM that characterizes a particular environment at a particular speaking rate. In the online phase with the speech rate training system prepared in the offline phase the super vector for a new testing environment at a new speaking rate is estimated based on a stochastic matching criterion. This paper focuses on a method for enhancing the coverage and construction of speech rate training at different speech rate in offline phase. The proposed Speech rate training framework was evaluated on the aurora2 connected digit recognition task.