Pitch extraction of esophageal speech

Abstract

A number of studies have been devoted to the description of the prosodic characteristics of esophageal speakers. However, pitch extraction of esophageal speech is difficult because of its low pitch, of the often quasi-periodic character of the vibration of the pharyngo-esophageal segment, and because stoma noise may be added to the signal. Commercially available hardware F0 extractors often produce artefacts in the analysis of esophageal speech. Consequently it may be asked whether the harmonic source model of laryngeal speech is inappropriate for esophageal speech, or whether it simply constitutes an extreme of this model in terms of the proportion of noise it contains. In a number of recent studies, cepstral analysis has been applied to pitch extraction of esophageal speech. This procedure handles low fundamentals well, but is computationally slow and complex. In this study, several other pitch extraction algorithms which are computationally simpler than the cepstrum will be compared to cepstral analysis in regard to their adequacy for F0 extraction with esophageal speech. Some of these, especially the harmonic product spectrum and the harmonic sum spectrum, have been shown to be superior to the cepstrum in the analysis of noise speech signals. [Work supported by the Norris Cancer Hospital, USC.]