Unfolding speaker clustering potential: a biomimetic approach
MM '09 Proceedings of the 17th ACM international conference on Multimedia
| Hi-index | 0.00 |
Psychoacoustic experimentation shows that formant frequency shifts can give rise to more significant changes in phonetic vowel timber than differences in overall level, bandwidth, spectral tilt, and formant amplitudes. Carlson and Granstrom's perceptual and computational findings suggest that, in addition to spectral representations, the human ear uses temporal information on formant periodicities ('Dominant Frequencies') in building vowel timber percepts. The availability of such temporal coding in the cat's auditory nerve fibers has been demonstrated in numerous physiological investigations undertaken during recent decades. In this paper we explore, and provide further support for, the Dominant Frequency hypothesis using KONVERT, a computational auditory model. KONVERT provides auditory excitation patterns for vowels by performing a critical-band analysis. It simulates phase locking in auditory neurons and outputs DF histograms. The modeling supports the assumption that listeners judge phonetic distance among vowels on the basis formant frequency differences as determined primarily by a time-based analysis. However, when instructed to judge psychophysical distance among vowels, they can also use spectral differences such as formant bandwidth, formant amplitudes and spectral tilt. Although there has been considerable debate among psychoacousticians about the functional role of phase locking in monaural hearing, the present research suggests that detailed temporal information may nonetheless play a significant role in speech perception.