Effects of wavelet compression of speech on its Mel-Cepstral coefficients
Katrina L. Neville, Hussain, Zahir · RMIT Research Repository (RMIT University Library) · 2009
This work looks at the compression of speech data using wavelet compression techniques and the effect such compression has on the speech data. Two wavelet compression techniques are utilised, these are: thresholding, where small coefficients in a wavelet decomposition are set to zero and the second method is using low-subband filtering of the coefficients. Two methods of error analysis are also used to determine the quantitative effect these compression methods have on the speech. Firstly the error is determined from the straight compressed and reconstructed speech and secondly the error is determined from the Mel-Frequency Cepstral Coefficients (MFCCs) which are features used in speech and voice recognition. The results of this work show that thresholding the wavelet coefficients causes a relatively small amount of error to occur in the straight reconstructed compressed speech but a great deal of error in the extracted MFCCs while the low-subband filtering method gave smaller error for the extracted MFCCs but greater error for the actual speech.