Two novel FDLP based feature extraction methods for improvement of speech recognition

Abstract

In conventional automatic speech recognition systems, linguistic information of the speech signal are usually acquired from short-time frames about 10–30 ms. In this paper we have proposed two novel methods extracting the long-term information of the speech signal. Both of the methods are based on “sub-band FDLP” which divides the long-time frame of signal into several sub-bands. Using the MFCC algorithm, we are able to represent the long-term temporal features of the each sub-band. Our results show that the proposed methods could improve the recognition ratio by %1.73. The proposed methods were evaluated using the FarsDat database and the method's robustness against different conditions of noise was experimented.