One problem is finding the times of all the excitation pulses driving a speech synthesizer. Ken Stevens at MIT hypothesized that the vocal quality problems of LPC-synthesized speech are readily solvable by updating coefficients with each pitch period, but an effective automatic system for labeling each pitch period is the rub. For laboratory speech perception experiments, Stevens proposed hand labeling.
Matched-filter theory is offered as the rationale for the widely-used autocorrelation pitch tracker. This pitch tracker, however, uses one pitch period as the template to find the neighbor, and thus does not provide the absolute registration of pitch events. A matched filter template based on an idealization of the vocal fold closure pulse has proved to be effective in automatic labeling the correct registration of vocal fold closure events for both adult male and female voices. The embarrassingly-simple realization of this filter consists of running the LPC residual signal through a 2nd-order digital filter and picking peaks. Another problem is tracking the positions of gold pellets placed on the tongue to monitor speech movement. The problem is locating a pellet center within a raster, subject to unknown background absorption in bone and fillings and a low photon count. The maximum-likelihood solution for Poisson photon statistics and an assumption of unknown but constant background absorption results in a matched filter template consisting of the log of pellet absorption.
Time and Place: Wed., Oct. 13, 3:30-4:30 pm in 4610 Engr. Hall.
SYSTEMS SEMINAR WEB PAGE: http://www.cae.wisc.edu/~gubner/seminar/