Wavelet based Cepstral Coefficients for neural network speech recognition

Traditional cepstral analysis methods are often used as part of feature extraction process in speech recognition. However the cepstral analysis method uses the Discrete Fourier Transform (DFT) in one of its computation process. The DFT uses fixed frame resolution to analyze frames of signal thus it will result in an analysis that would not accurately analyze localized events. This paper investigates the use of the Discrete Wavelet Transform (DWT) for calculating the cepstrum coefficients. Two wavelet types with different decomposition level are experimented to yield the cepstrum which is called the Wavelet Cepstral Coefficient (WCC). To test the WCC speech recognizing task of recognizing 26 English alphabets were conducted. Under same number of feature dimension the WCC outperformed the MFCC with about 20% in terms of recognition rate under both speaker dependent and speaker independent task.