Noise Robust Speech Features for Automatic Continuous Speech Recognition using Running Spectrum Analysis

Abstract

In this report, new robust speech feature is introduced and applied for an automatic continuous speech recognition system. Using these features, the noise robust continuous speech recognition can be realized. The new running spectrum analysis (RSA) method is used in order to remove un-speech components over 15 Hz in modulation spectrum domain. Using RSA, speech features are emphasized for the design of tri-phone HMM where the tri-phone HMM is used in continuous speech recognition. In order to show the performance of the developed system, some comparisons with conventional one are given in experiments.