Uso de reconhecedor de fala para identificar fonemas sustentados em fala normal e disfônica

Abstract

This work aims to apply speech recognition technology through probabilistic evaluation, analyzing the hit rate and accuracy in the identification of sustained phonemes under different vocal conditions. The focus is directed at individuals aged between 45 and 60 years, divided into groups with normal speech and dysphonic speech due to Reinke’s edema. The manual labeling of speech signals was performed using the Praat software, while the recognition process was conducted through Hidden Markov Models (HMM), implemented in the HTK (Hidden Markov Model Toolkit) tool. The research aims not only to evaluate the feasibility of using the speech recognizer in a clinical context, but also to contribute to the development of diagnostic support tools in speech-language pathology. The results obtained showed a hit rate of 89.54% and an accuracy of 65.09%, indicating that, despite the moderate accuracy, the method has the potential for optimization with the use of more robust databases and adjustments to the model parameters.