ICA 2010 paper

There are several diseases that affect the human voice quality which can be organic or neurological. Acoustic analysis of voice features can be used as a complementary and noninvasive tool for the diagnosis of laryngeal pathologies. The degree of reliability and effectiveness of the discriminating process depends on the appropriate acoustic feature extraction. This work presents a parametric method based on cepstral features to discriminate pathological voices of speakers affected by vocal fold edema and paralysis from healthy voices. Cepstral, weighted cepstral, delta cepstral, and weighted delta cepstral coefficients are obtained from speech signals. A Vector Quantization is carried out individually for each feature in the classification process, associated with a distortion measurement. The goal is to evaluate a performance of a classifier based on the individual and combined cepstral features. The average, the product and the weighted average are the different combination strategies applied yielding a multiple classifier that is more efficient than each individual technique. To assess the accuracy of the system, 153 speech files of sustained vowel /ah/ (53 healthy, 44 vocal fold edema and 56 paralysis) of the Disordered Voice Database from Massachusetts Eye and Ear Infirmary (MEEI) are used. Results show that the employed parameters are complementary and they can be used to detect vocal disorders caused by the presence of vocal fold pathologies.