Enhancing Robustness of Speech Recognition by Approach of Feature with Confident Weight

L. Ge, K. Shirai (Japan), and Y. Ge (PRC)


Natural Language Processing, Robust


Enhancement of robustness has become one of research focuses of acoustic speech recognition system. In recent works, Missing Feature Theory (MFT) has been proved an available and considerable solution for robust speech rec ognition based on either ignoring or compensating the un reliable components of feature vectors corrupted mainly by band-limited background noise. Because of MFA classify ing in binary way and necessarily of dealing with the cep stral feature, this paper proposes three new approaches based on confidence analysis. Approach of Feature with Confident Weight(AFCW) estimates the confidence of each feature component as its weight and describes the effect of noise in a more precise way. The other two approaches, SC(Simple Cepstral)- and TC(Total Cepstral)-AFCW, can be regarded as AFCW on cepstral domain. Experimental results show proposed approaches could improve the rec ognition accuracy significantly in adverse environment, including stationary and non-stationary noise environments. The paper is organized in the following. In Section 2, We first present basic AFCW algorithm which is based on spectral subband confidence analysis, and how it relates to missing feature theory. In the next section we discuss the implementation of the algorithm in cepstral domain. Then in the final two sections we present our experimental re sults and finally conclude the paper with some discussion on potential directions for further research on the subject.

Important Links:

Go Back