Speech Recognition Using Sparse Discrete Wavelet Decomposition Feature Extraction

In this paper, a new feature extraction algorithm for speech recognition using sparse discrete wavelet decomposition (SDWD) is proposed. The recognition system contains the following stages: speech data acquisition and preprocessing, speech signal decomposition using the SDWD, feature extraction, and artificial neural network (ANN) classifier. The task of the developed SDWD is to decompose speech signal into band signals based on on the Mel filter bank frequency specifications. Similar to the Mel frequency cepstral coefficient (MFCC) method, the logarithmic values of the filter bank energies are computed and then a discrete cosine transform (DCT) is applied to these logarithmic values to extract the feature. Our experimental results using the ANN classifier demonstrate that our proposed SDWD feature extraction algorithm outperforms over the MFCC and discrete wavelet packet transform (DWPT) algorithms.