Comparison of two frequency domain pitch detectors based on phonetic performance evaluation

A comparison of two pitch detection algorithms using frequency domain analysis is presented. The algorithm showing the best results based on a perfomance evaluation done by a phonetician will be selected for use in an automatic speech recogni tion (ASR-) system based on the acoustic phonetic approach using knowledge-bases with an expert system. The first algori thm is the Seneff pi tch detection algori thm, which uses purely harmonic information taken from a preselected part of the amplitude FFT-spectrum. The harmonics are picked from this spectrum and peak processing heuristics estimates the pi tch. The second algori thm is the Charpentier pi tch detection algori thm, which uses information from the ampli tude and the phase spectrum. Based on a combination of the ampli tude and the phase information the pitch is estimated. The paper will compare the resul ts from the two algori thms based on evaluation performed by a phonetician. The most reliable pi tch algorithm will be used in an ASR-system, which is based on the acoustic phonetic method. The system is running in a microcomputer environment, where phonetic feature extraction (formants, pitch etc. ) is performed in the low level of the system in real time preprocessors and used in the high level of the system, which is a dedicated knowledge-based expert system. The pi tch algori thm will be implemented in a floating-point 32 bit signal processor. Aalborg University, Speech Technology Centre, 19, Strandvejen, DK 9000 Aalborg, Denmark