Pitch detection by data reduction

This paper presents an algorithm that determines the fundamental frequency of sampled speech by segmenting the signal into pitch periods. Segmentation is achieved by identifying those samples of the waveform corresponding to the beginning of each pitch period. The segmentation is accomplished in three phases. First, using zero crossing and energy measurements, a data structure is constructed from the speech samples. This structure contains candidates for pitch period markers. Next, the number of candidate markers within this structure is reduced utilizing syllabic segmentation, coarse pitch frequency estimations, and discrimination functions. Finally, the remaining pitch period markers are corrected, compensating for errors introduced by the data reduction process. This algorithm processes both male and female speech, provides a voiced-unvoiced decision, and operates in real time on a medium speed, general purpose computer.