Technical Considerations in Computation of Spectral Harmonics-to-Noise Ratios for Sustained Vowels

Abstract

This paper explores technical issues affecting computed measures of the relative level of noise in the frequency spectrum of a vowel. This type of measure has been proposed for quantification of hoarseness in pathological speakers. An analysis of synthesized vowels was used to test the influence of vowel type, fundamental frequency, perturbation type, perturbation level and quantization. The algorithms were shown to be highly sensitive to errors in pitch-period demarcation, and a dependency on jitter perturbations, fundamental frequency, and vowel type was demonstrated. Relationships between algorithm performance and methods of spectrum estimation were discussed, and approaches for reducing the dependencies were proposed. Finally, a method for achieving a significant reduction in computation time was described.