Resonant bandwidth estimation of vowels using clustered-line spectrum modeling for pressure speech waveforms

The estimation of resonant frequency bandwidths is a fundamental issue related to the quality of spoken vowels and vocal-tract acoustics. In this article, we discuss our analysis of bandwidths using clustered line-spectrum modeling (CLSM) of the pressure waveforms of vowels on a cycle-by-cycle basis with reference to Lx waveforms from an electrolaryngograph recorded at the same time as the speech signal. We used CLSM to decompose the waveforms into three dominant resonant (modal) oscillations with almost exponentially decaying envelopes. The modal (so-called formant) frequencies were observed in a wide frequency range from 100 (Hz) to over 4 (kHz). The modal bandwidths were estimated from the decaying constants of the modal oscillations and were wider than those reported in the literature under the closed glottis condition. The bandwidths increased for both male and female speakers as the formant frequencies became higher. The bandwidths for females, however, were wider with greater variances than those for males. We could effectively represent a cycle of a vowel record shorter than 10 (ms) by CLSM based on the least squares error criterion in the frequency domain. We thus confirmed that cycle-by-cycle analysis using CLSM is a practical approach to characterizing vowel sounds in terms of dominant frequencies using their modal bandwidths.