An insight to the automatic categorization of speakers according to sex and its application to the detection of voice pathologies: A comparative study

Abstract

An automatic categorization of the speakers according to their sex improves the performance of an automatic detector of voice pathologies. This is grounded on findings demonstrating perceptual, acoustical and anatomical differences in males' and females' voices. In particular, this paper follows two objectives: 1) to design a system which automatically discriminates the sex of a speaker when using normophonic and pathological speech, 2) to study the influence that this sex detector has on the accuracy of a further voice pathology detector. The parameterization of the automatic sex detector relies on MFCC applied to speech; and MFCC applied to glottal waveforms plus parameters modeling the vocal tract. The glottal waveforms are extracted from speech via iterative lattice inverse filters. Regarding the pathology detector, a MFCC parameterization is applied to speech signals. Classification, in both sex and pathology detectors, is carried out using state of the art techniques based on universal background models. Experiments are performed in the Saarbrücken database, employing the sustained phonation of vowel /a/. Results indicate that the sex of the speaker may be discriminated automatically using normophonic and pathological speech, obtaining accuracy up to 95%. Moreover, including the a-priori information about the sex of the speaker produces an absolute performance improvement in EER of about 2% on pathology detection tasks. RESUMEN:Una categorización automática de los hablantes de acuerdo con su sexo mejora el rendimiento de un detector automático de patologías de voz. Esto se fundamenta en hallazgos que demuestran diferencias perceptuales, acústicas y anatómicas en voces masculinas y femeninas. En particular, este trabajo persigue dos objetivos: 1) diseñar un sistema que discrimine automáticamente el sexo de hablantes utilizando habla normofónica y patológica, 2) estudiar la influencia que este detector de sexo tiene sobre el acierto de un posterior detector de patologías de voz. La parametrización del detector automático de sexo se basa en MFCC aplicados sobre señales de voz; y MFCC aplicados a formas de onda glotal junto a parámetros que modelan el tracto vocal. Las formas de onda glotal se extraen de la voz a través de un filtrado inverso iterativo en celosía. En cuanto al detector de patologías, una parametrización MFCC se aplica a señales de voz. La clasificación, tanto en los detectores de sexo como de patología, se lleva a cabo con técnicas del estado del arte basadas en modelos de base universal. Experimentos son realizados sobre la base de datos Saarbrücken empleando la fonación sostenida de la vocal /a/. Los resultados indican que el sexo del hablante puede ser discriminado automáticamente utilizando habla normofónica y

On the design of automatic voice condition analysis systems. Part II: Review of speaker recognition techniques and study on the effects of different variability factors

“…Results indicate that a segmentation according to the sex of the speaker improves the performance, which in absolute terms varies between 0.3% to 4% depending on the dataset. These results are in line with those found in literature, where the classification accuracy of an automatic detector of pathology is lightly improved by using a manual segmentation of the dataset according to the sex of the speakers [46]. Results suggest that partitioning the dataset according to this extralinguistic criterion improves performance, indicating the usefulness of hierarchical systems that decompose the voice pathology detection problem into smaller subproblems.…”