Skip to main content

Research Repository

Advanced Search

Efficient digital techniques for speech processing

Efficient digital techniques for speech processing Thumbnail


Abstract

Computationally efficient digital signal processing algorithms suited for speech signals are investigated, A new efficient time domain algorithm for estimating the pitch period of voiced speech is presented.
This algorithm has no multiply operations and can be implemented in integer arithmetic without scaling on a 16-bit microprocessor, The algorithm gives a low error rate with signal to noise ratio higher than 10 dB, Moreover, a good signal intensity estimation is obtained as a by-product of the algorithm.
The importance of the zero-crossing counts of a differentiated speech waveform is explored in terms of a discrete mathematical analysis.
The potential of this parameter is shown by its use in a new speaker verification system, The verification score obtained using this parameter in combination with the intensity compares well with the score obtained using only the pitch period parameter. These three parameters have also been compared in terms of their ability to discriminate between speakers,
The computational effort necessary to extract the zero-crossing count of differentiated speech is very small and it can be extracted using a microprocessor in. real time,
An efficient way of creating reference templates using a nonlinear mapping technique to cater for intraspeaker variations is presented. Results show that the speaker verification score is improved when intraspeaker variations are considered in creating reference templates,
A speaker dependent digit recognition system has been implemented using Burg’s Partial Correlation coefficients and their nonlinear transforms, The results show that the recognition score obtained is 100 per cent with three or more Burg's coefficients, and that a simple 'city block’ distance measure is adequate,
Finally a new computationally efficient multiplication technique which speeds multiplication at the expense of memory space is developed.

Publicly Available Date Mar 29, 2024

Files




Downloadable Citations