| The application of the wavelet transform to the generation of feature vectors for a closed set, text independent speaker identification problem is examined. Three sets of wavelet derived feature vectors are applied to the speaker identification problem and the identification accuracy is compared against the accuracy obtained using feature vectors generated from standard cepstral coefficients.; Results obtained indicate that the half band filter bank structure of the standard wavelet transform is too restrictive for the analysis of speech signals in speaker identification applications, since it does not provide the necessary spectral resolution in the range 0-1.0 kHz. Standard signal processing techniques based on the Fast Fourier Transform, while involving slightly more operations, do not have this limitation and therefore seem to be more suitable for the processing of speech signals for speaker identification purposes. |
