| Abstract |
Voice is our primary communication tool. This human instrument is important for many professionals such as singers and teachers. With the appearance of various voice disorders, clinical studies dictate the necessity of voice function assesment. Investigation techniques seek to provide for comprehensive information of the voice production mechanism and more specifically, the examination of vocal-folds vibratory behavior and the corresponding acoustiv features.
The vocal folds are composed of twin infoldings of mucous membrane stretched horizontally across the larynx. Their vibration modulates the air flow expelled from the lungs during phonation, thus functioning as an acoustic excitation, i.e. the source of voiced speech. Electroglottography is a non-invasive investigation technique which measures the impedance related to the glottal area, that is the area between the vocal cords. Direct observation of the vocal-fold vibrations can only be achieved with invasive imaging techniques, by recording the vocal-fold vibrations from a top-view position in the larynx.
High-speed videoendoscopy is the most promising approach to directly assess vocal-fold vibrations. Automatic segmentation of the glottal area from the acquired data is a major challenge. In this work, we address this problem by proposing a local-based active contours framework for glottis segmentation. The method provides localizing of the glottal area with high accuracy and allows the contour to split and merge, according to the local-fold vibratory pattern. The method is fully automatic and parameter selevtion is performed automatically, based on the sequence statistics and empirial observations performed on a wide range of high-speed sequences.
Manual verification resulted on average in ±10 pixel modifications, which correspond to less than 1% of the average glottal area. The vast amount of data that has to be evaluated both quantitatively and qualitatively dictate the need for dimensionality reduction of the spatio-temporal information and efficient represenation of the high-speed data in a compact, handy and lossless way. One and two dimensional represenations have been used for that reason. A comparison has been conducted on glottal parameters (fundamental frequency, open quotient) estimated on the electroglottogrpahic signal and the extracted glottal area signal, using the electroglottographic signal as reference.
|
Glottal source analysis : a combinatory study using high-speed videoendoscopy and electroglottogrpahy
