Perceptual relevance of source spectral slope measures

Abstract

Researchers broadly agree that the spectral slope of the voice source is an important concomitant of voice quality. Many measures of source spectral slope have been proposed, including the relative amplitudes of the lowest few harmonics, the ratio of energy in low- versus high-frequency bands, and the average deviation from an ideal slope. It is unclear which (if any) of these measures best reflects the differences in vocal quality that result from the underlying acoustic variability. To examine this issue, a large corpus of voice samples was inverse filtered, and spectra were calculated for resulting source pulses. Different measures of spectral slope were calculated for each voice, and correlations among measures were examined. Finally, several series of synthetic stimuli were created in which only the source spectral slope varied in steps. Listeners judged the similarity of stimuli within each series. Similarity responses were evaluated with multidimensional scaling, and the resulting perceptual spaces were interpreted in terms of the different measures of source spectral slope. Measures that are highly correlated with the perceptual spaces reflect perceptually important aspects of the source signal. [Research supported by NIDCD.]