| Objective methods for evaluation of pathological voices have long been sought for both clinical use and basic research; classification of types and severity of voice defects has been largely accomplished by trained clinicians using subjective rating systems. Automatic modeling and parameterization of pathological vowel samples could provide a useful advance in this effort. Towards this goal, this work develops approaches to automatically analyze and parameterize voices with methods specialized to pathological vowels. The nonperiodic components of sustained vowels, which are prominent features of pathological voices, are modeled as three phenomena: frequency modulation (FM), amplitude modulation (AM), and aspiration noise. Improved modeling accuracy was obtained by AM and FM demodulation of the original voice prior to aspiration noise analysis. Pathological voices were re-synthesized incorporating the models and measured levels of nonperiodic components. The resulting waveforms sounded more natural than the synthetic voices that did not model AM/FM and/or aspiration noise. Major results of the work include a real-time vowel synthesizer allowing instantaneous adjustment of parameters, a flexible software vowel synthesizer, and application of an external stimulation technique to aid in the identification of the vocal tract transfer function in the case of voices with spectrally deficient glottal driving waveforms. |
