Measuring voice quality in dynamic speech

Abstract

Different voice qualities are used in normal speech to convey phonological contrasts (e.g., Zapotec, Gujarati, etc.), prosodic information, emotions, etc. When measuring changes in voice quality in dynamic speech, it is necessary to use a technique that can capture small, relative differences under quickly changing conditions. Although a number of techniques are available for assessing voice quality (e.g., perceptual assessment, qualitative assessment of the time amplitude waveform, qualitative assessment of the spectrogram, quantitative measurement of the spectrum, and quantitative measurement of the voicing source), they have, for the most part, been designed for capturing voice quality differences in sustained vowels or carefully matched speech conditions. This study develops a technique for relative voice quality measurement which ascertains voice quality differences across diverse words, pitches, and speakers. Specifically, inverse filtering is used to obtain the differentiated glottal flow for each vowel in the dynamic speech, which is theoretically robust across different vowel qualities and pitches. Then, for comparison, a baseline voice quality measurement is obtained for each word from each speaker to control for segmental and personal voice quality differences. Thus, the relative voice quality measurement will be resistant to the variations in dynamic speech. [Work supported by NIH and NIH/NIDCD.]