Libro de Resúmenes / Book of Abstracts (Español/English)

Resumenes 163
diferencial automático que puede resultar de ayuda al profesional
fonoaudiólogo.
Automatic classification of pathological voices
Even if normal phonation and voice disorders can be qualitatively
distinguished by the well trained human ear, nowadays it does not exist any
system capable to perform a quantification and estimation of the
distribution of the data corresponding to different pathologies. This work is
motivated by the clinical interest made clear by the phonoaudiology
community and the need to develop new techniques that could orientate in
this sense.
In phonoaudiology it is clinically interesting to determine normal for
pathological phonations. This task results more complicated when it is
necessary to precise the pathology, given that in certain cases the voices of
the patients can have similar characteristics but different pathologies. An
important case is to discriminate between two dysphonias: muscular tension
dysphonia (MTD) and spasmodic dysphonia.
This importance lies in that the treatments are totally different, and only a
few highly trained specialists can correctly diagnose the pathology. Classic
temporal measures based on the fundamental frequency and consecutive
periods amplitude variations help to distinguish between normal and
pathological voices, but they are not enough to discriminate between the
dysphonias above mentioned. At the present time there are no tools that
allow the automatic classification of these pathologies. This absence motives
us to develop new techniques that help in the automatic diagnosis from a
sustained vowel /a/ record.
In the present work, we first classify the data in normal and
pathological voices. Next, we proceed to discriminate between the two
pathologies of our interest. With this aim in view we use different
parameters calculated from the voice signal. The parameters are: the
fraction of locally unvoiced signal, degree of vocal breaks, three different
measurements of jitter (degree of difference between consecutive
fundamental frequencies of sustained voice signal), and three different
measurements of shimmer (degree of difference between the amplitudes of
consecutive periods). With these parameters we build the patterns
representing each patient. Neural networks (NN) and support vector
machines (SVM) performances as automatic pattern recognizers are
analyzed. Neural networks are trained by different methods, obtaining
results that depend on such methods. In case of SVMs, its performance for
different kernels is studied here. The presented advances point out a way
for developing an automatic differential diagnosis method that could help to
the phonoaudiologist.