Colored L-l filters and their application in speech pitch detection

This paper develops colored L-l filters with tap bias and evaluates their performance in nonlinear system modeling and speech pitch detection. L-l filters are nonlinear weighted sum based filters that jointly exploit the temporal and rank orderings of observation samples. The use of joint ordering allows L-l filters to effectively process nonstationary signals with heavy tailed distributions. Utilizing this method of incorporating temporal and rank order information, however, requires N/sup 2/ parameters to describe a window size N filter. This polynomial growth in the number of filter parameters limits the use of L-l filters to applications that can be satisfactorily addressed with moderate sized windows. It is shown here that performance increases in temporal-rank order filters can often be achieved by expanding the window size. A method for reducing the L-l filter parameter space, and thus relaxing the window size limitations, is developed through the use of group theory based coloring, or quantization, of rank order information. Additionally, the modeling ability of colored L-l filters is improved through the simple inclusion of tap biases. Optimization procedures are developed for determining filter tap weights and biases as well as the rank order coloring. It is shown that the window size/rank order quantization tradeoff has advantages in many applications and that the inclusion of tap bias results in improved nonlinear system modeling. Finally, a method for utilizing colored L-l filters in the determination of glottal closure instants (GCIs) for speech pitch period identification is developed. Results presented show that this method is a more accurate GCI indicator than a commonly used wavelet approach and requires less processing time.