Pitch detection using slope overload noise in an adaptive delta modulation process

Abstract

This algorithm uses the slope overload noise or error introduced by the adaptative delta modulation (ADM) coding process. This error, greatest at a pitch pulse, decreases in magnitude during a pitch period and displays distinctive autocorrelation differences between voiced and unvoiced speech segments. In this application, the step adaptation and maximum step size parameters are selected to maximize the slope overload error. Autocorrelation techniques applied to the error signal do not precisely delineate the transition points between voiced and unvoiced segments in the time waveform. Hence several other features of the error signal are incorporated into the algorithm. These are the absolute correlation value, normalized peak correlation value, and percent variation of pitch period in adjacent frames. The exact position of the pitch pulse is determined by searching for the maximum value in the error waveform within a variable size time window. Current experimental efforts are directed towards further validation of the preliminary results. This technique of pitch detection is compared with that of the simplified inverse filtering technique (SIFT), reported by Markel [IEEE Trans. Audio Electroacoust. AU‐20(5), 367–377 (Dec. 1972)l. Simplicity of the ADM process over the linear predictor approximation used in the SIFT algorithm makes this new technique more attractive to hardware implementation.