Desarrollo de una herramienta computacional para la evaluación de la calidad de voz con base en la escala GRBAS

Abstract

Voice quality analysis has become a routine activity in clinics and hospitals, where they are performed by voice professionals (speech therapists); These analyses are generally performed based on the GRBAS scale, and present subjective characteristics highly influenced by experience, level of education of staff, among others (Gordillo, 2018). For this project, the database of synthetic voices developed by the Antonio Nariño University was implemented through the Evaper application, with the purpose of developing a computational tool for the evaluation of voice quality, through the extraction of both acoustic and statistical vocal characteristics, and the implementation of machine learning systems to give a respective diagnosis based on the GRBAS scale. As a first result, it was found that the algorithm in charge of performing the extraction of vocal features presented a strong level of correlation with respect to the Praat software, software that was considered as a standard system due to its very significant trajectory in the field of speech therapy; reaching a Spearman rho correlation > 0.85, thus validating the algorithm dedicated to the extraction of features, and the implemented methodological process. As a second result, it was found that the classification models implemented in this project presented a high level of accuracy, with the exception of one of the parameters of the female gender (Roughness), due to the existence of an error in the database, since it presented a lack of information for this gender; the results obtained in percentage scale of the level of accuracy of the models that make up the tool are: Hoarseness Models for male gender =71.2% , Tension Asthenia Models for male gender = 84.8% , Tension Asthenia Models for male gender = 93.3%, Grade Models for male gender = 93.6%, Hoarseness Models for female gender = 40.5% , Tension Asthenia Models for female gender = 90.5% , Tension Asthenia Models for female gender = 97.7%, Grade Models for female gender = 95.3%. As a last result, after performing a preliminary preview of phase 2, which consisted of evaluating the performance of the tool implemented 10 real voices provided by the Universidad del Valle, low levels of accuracy and correlation were obtained, being these values equal to: Accuracy <= 50% and -0,2 < Correlation <= 0,5. The results obtained through the evaluation of the models developed with synthetic voices, allowed validating the performance of the computational tool, however, the results after performing the preliminary preview, despite not being very significant due to the low amount of data, showed that it is necessary to carry out an analysis of the tool in order to make the appropriate modifications to improve the effectiveness of its operation when implemented with real voices.