Designing of speaker based wavelet filter

Speech recognition emphasis on capturing typical properties of phonemes, irrespective of any speaker-originated variations in speaking style and pronunciation, while speaker recognition aims at exploiting these individualities. These different traits of the speaker recognition tasks compelled us to search a more general approach. The main motto of this paper is to study speech style that would favor the speaker recognition process, and provide a better categorization of the individual voices. A unique filter design process has been proposed for individual speaker. The wavelet system that is developed can be used for analysis and synthesis of the speech signals. The performance analysis of wavelet filter coefficients and maximally flat wavelet filter is given on the basis of performance parameters mentioned above. Estimation of filter coefficients is done to analyze the estimation of analysis wavelet filter from a given signal. After the estimation of filter coefficient, delay is provided by tuning the filter parameter at each coefficient and modified filter coefficient is used to reconstruct the orthogonal wavelet. This is validated by calculating, the energy compaction ratio (ECR), the absolute mean squared errors (AMSE) and minimum mean square error (MMSE) between the two signals (original signal and reconstructed signal). The designed wavelet outperforms the predefined wavelet in terms of three performance parameters and the comparison between categorization rates of the two wavelets are discussed in the result section.