[Diagnosing voice disorders in children with cochlear implantation and hearing aids using artificial intelligence systems]

ABSTRACT

The purpose of this study is to quantify the voice disorders in children with cochlear implantation and hearing aids. Until now, quantifying voice disorders has been done subjectively by speech experts and it is for the first time that the preset study tends to run an objective experiment using signal processing features. 4 levels were considered to be qualify speech. Linear and nonlinear features were extracted from 5 Farsi words "mashin', "mar', "moosh', "gav" and "mowz" uttered by 30 subjects and then put into hidden Markov classifiers. Classifier outputs then were fused together to have better accuracy. The main hypothesis of the study is to answer this question Can we separate children into 4 levels based on their voice features? Voice features including "fundamental frequency, first formant, second formant, third formant, first to second formant ratio, third to second formant ratio, Rational Intensity, nasality, approximate entropy and fractal dimension were extracted from speech segments and then are were given to artificial decision making system [classifiers]. The results show that classifiers can separate 4 levels of voice disorders with the accuracy of 93.1%. Among the introduced features, first to second formant ratio and third to second formant ratio can be used directly to track voice recovery after using cochlear implantation or hearing aid. The output of this research study can act as a speaker independent system to help speech specialists with evaluating voice disorder recovery in children who fall in the same range of age