A geometric approach to real time pitch detection

An approach to time-domain pitch detection based on the concepts of center of mass is presented and evaluated. Excursions, or bumps, in speech signals are treated as geometric areas and replaced by their total mass lumped at the centers of mass. Intervals between masses are grouped into candidate classes. Coincidence and coherence indices of these classes are computed to determine the most likely pitch estimate. Postprocessing consists of a simple error-correction and silence-detection scheme. This algorithm compares favourably in performance with the autocorrelation method, using pitch contours from electroglottograph signals as a reference. The algorithm is tested in noisy environments simulated by uniformly distributed white noise and multitalker babble noise. Results show that the algorithm is robust and accurate. The implementation of this algorithm for a vibrotactile device to aid lipreading is described