A powerful post‐processing algorithm for time domain pitch trackers

Existing pitch tracking algorithms are not precise nor reliable enough to be useful in feature‐based recognition systems. It is possible, however, to analyze the errors produced by a particular algorithm and then reduce these errors using post‐processing techniques. Error patterns were analyzed for a time‐domain pitch tracker [W. H. Tucker and R. H. T. Bates, IEEE Trans. ASSP‐26, 597–604 (1978)] and a post‐processing algorithm, based on artificial intelligence techniques, was written in order to eliminate errors. Performance was compared for the original and modified versions of the pitch tracker for a number of speakers using both isolated words and sentences. All types of errors were reduced by the post‐processing algorithm. Voiced‐voiceless decisions were performed with less than 1% error for 2080 letters produced by males and females. The fundamental frequency microstructure was tracked sufficiently well to be used in extracting phonetic features in a feature‐based recognition system.