The design of voice pathology automatic detection systems is gaining attention in the last few years for presenting advantages compared to traditional diagnosis methods. However, the performance of these systems is influenced by aspects related to the inter-speaker variability, and specially to the heterogeneity introduce by gender differences. To overcome that, a gender recognizer may be employed as a preprocessing stage in order to stratify the speakers and further adjust the detectors to the specific characteristics of each target group. Nevertheless, the reliability of gender recognizers on pathological speech has not been investigated. Having this in mind, the present paper studies the effectiveness of an automatic gender recognizer, based on mel frequency cepstral coefficients and gaussian mixture models, on normal and pathological speech. The analysis is carried out parameterizing the speech, the glottal waveform extracted from speech via inverse filtering, and a vocal tract model. The experiments were carried out using sustained vowels taken from the Saarbrucken and UPM voice disorders databases, and suggest that the gender might be effectively classified when using the proposed methodology. They also suggest that gender recognizers can be successfully employed as a preprocessing stage for a more accurate design of gender-dependent pathology detection systems.
Bibliographic reference. Gómez-García, J. A. / Godino-Llorente, Juan Ignacio / Castellanos-Domínguez, G. (2013): "Automatic gender recognition in normal and pathological speech", In INTERSPEECH-2013, 1707-1711.