4aSC18. Optimization for source waveform synthesis of pathological voices.

Session: Thursday Morning, May 16

Author: Brian C. Gabelman
Author: Jody Kreiman
Author: Bruce Gerratt
Location: Voice Lab., Div. of Head/Neck Surgery, UCLA, 31-24 Rehab. Ctr., Los Angeles, CA 90095-1794
Author: Abeer Alwan
Location: UCLA, Los Angeles, CA 90095-1794


This paper describes an algorithm that fits analytical source models of the glottal volume velocity to glottal source waveforms from inverse-filtered vowels. The analytical models include the KGLOTT88, LF, and simplified LF [Qi and Bi, J. Acoust. Soc. Am. 96, 1182--1185 (1994)]. In the first stage of the algorithm, curve features (peak values, zero crossing, period, etc.) are automatically extracted from samples of pathological voices and least-squares fitting is performed. The parameters are then fine-tuned based on the perceptual quality of the synthesized voices. Goodness of approximation of the resultant source function is determined by perceptually comparing the final synthesized vowel sounds to the original sound. Limitations of current source models with respect to modeling pathological voice will be discussed. [Work supported by NIDCD.]

from ASA 131st Meeting, Indianapolis, May 1996