Conversão de fala sussurrada para fala normal usando modelos neurais

Abstract

Whispering is a common and secondary speech mechanism for communicating, however, in some cases, it can also be the main form to communicate, such as cases of people with pathologies of the larynx or who have suffered some type of damage to the vocal folds. Differences in characteristics between normal voice and whispered discussions have raised in speech recognition area, due to the difficulty of converting whispered to normal speech. This work presents the study of the characteristics of normal and whispered speech, and the implementation of the whispered to normal voice conversion system using multilayer perceptron networks and generative adversarial networks. The database used came from the partnership with Faculdade de Porto, which are vowel signs used in European Portuguese. To validate the study, three metrics were analyzed: Mel-Cepstrum Distortion, square root of the mean error of log(F0) and accuracy of the vowel classifier.