In this paper, we propose the use of the wavelet transform for the extraction of features for phonemes in order to overcome some of the shortcomings of short time Fourier transform. New log-energy based features are proposed using discrete wavelet transform as well as wavelet packets and their recognition performance has been evaluated. These features overcome the problem of shift variance as encountered in the features based on the discrete wavelet transform coefficients. The effect on the recognition performance by choosing different mother wavelets for the decomposition and window duration is also studied. Finally, a scheme based on the admissible wavelet packet has also been proposed and the results are discussed and compared with the frequently used Mel Frequency Cepstral Coefficients based features. The recognition performance of these features is further evaluated in the presence of different level of additive white Gaussian noise.
