DEVELOPING ‘STANDARD NOVEL ‘VAD’ TECHNIQUE’ AND ‘NOISE FREE SIGNALS’ FOR SPEECH AUDITORY BRAINSTEM RESPONSES FOR HUMAN SUBJECTS

In this research as a first step we have concentrated on collecting non - intra cortical EEG data of Brainstem Speech Evoked Potentials from human subjects in an Audiology Lab in University of Ottawa. The problems we have considered are the most advanced and most essential problems of interest in Auditory Neural Signal Processing area in the world: The first problem is the Voice Activity Detection (VAD) in Speech Auditory Brainstem Responses (ABR); The second problem is to identify the best De - noising techniq ue for Auditory Artifact removal in Speech ABR of Brainstem Speech Evoked Potentials. In VAD problem we have implemented Zero - Crossing Detection VAD, statistical algorithms (two algorithms) which are already a standard in Speech Processing VAD problems, an d then a third VAD we have presented is based on spectral subtraction method in which we have developed our own mathematical formula for the peak valley difference detection of the frequency spectra to detect the voice activity (we named it as SNRPVD VAD). These algorithms we applied on our data sets of EEG collected Brainstem Speech Evoked Potentials and compared their performances. VAD is verified and we found that SNRPVD VAD algorithm is working better than the Statistical VAD techniques and found to be it is detecting Voice even in more noisy data where statistical method could not detect. The second problem we considered is to de - noise the data from auditory artifacts and improve its SNR. We developed various De - noising techniques specifically: Yule - Wal ker Multiband filter; Cascaded “Yule - Walker and Comb” filter; Conventional Wavelets: Daubechies, Symlet, Coiflet Wavelet filtering; FAST Independent Component Analysis (FASTICA) filtering; Translation - Invariant (TI) Estimation filtering; Cycle Spin Transla tion Invariant Wavelets Independent Component Analysis (CSTIICA) Filtering approaches. We found the Wavelets are working better and TI wavelets are working far better than all; and then CSTIICA is working even better than TI wavelets. The performance meas ures considered are Signal to Noise Ratio (SNR) and Mean Square Error (MSE). Ultimately we observed that Wavelets are working surely as one of the best tools for de - noising neurological signals specifically Speech ABR signals. With these Novel observations in VAD and De - noising for Speech ABR, as both are relatively new and advanced areas in Audiology and as they gained wide interest in the last about a decade in scientific community of audiology scientists and engineers, Novel techniques are essentially em ergency and hence with our research we are contributing some advanced working ideas to the community of Auditory Neural Signal Processing.