A novel approach for speech feature extraction by Cubic-Log compression in MFCC

Abstract

Speech Pre-processing is measured as major step in development of feature vector extraction for an efficient Automatic Speech Recognition (ASR) system. A novel approach for speech feature extraction is by applying the Mel-frequency cepstral co-efficient (MFCC) algorithm using Cubic-Log compression instead of Logarithmic compression in MFCC. In proposed MFCC, the frequency axis is initially warped to the mel-scale which is roughly below 2 kHz and logarithmic above this point. Triangular filter are equally spaced in the mel-scale are applied on the warped spectrum. The result of the filters are compressed using Cubic-Log function and cepstral co-efficient are computed by applying DCT to obtain minimum MFCC feature vector for spoken words. These feature vectors are given as input to classification and Recognition phase. The system is trained and tested by generating MFCC feature vector for 600 isolated words, 256 connected words and 150 sentences in clear and noisy environment. Experiment results shows that with minimum MFCC feature vector is enough for speech recognition system to achieve high recognition rate and its performance is measured based on Mean Square Error (MSE) rate.