A powerful post-processing algorithm for time-domain pitch trackers

This paper describes a powerful post-processing algorithm for time-domain pitch trackers. On two successive passes, the post-processing algorithm eliminates errors produced during a first pass by a time-domain pitch tracker. During the second pass, incorrect pitch values are detected as "outliers" by computing the distribution of values over a sliding 80 msec window. During the third pass, remaining pitch pulses are used as anchor points to reconstruct the pitch train from the original waveform. The algorithm produced a decrease in the error rate from 21% obtained with the original time domain pitch tracker to 2% for words and sentences produced in an office environment by 3 male and 3 female talkers. In a noisy computer room errors decreased from 52% to 2.9% for the same stimuli produced by 2 male talkers. The fundamental frequency micro-structure is tracked sufficiently well to be used in extracting phonetic features in a feature-based recognition system.