Analysis and correcting pronunciation disorders based on artificial intelligence approach

The main aim of this study is to employ artificial intelligence and machine learning methods to assess and correct pronunciation disorders in post-traumatic military patients, acknowledging the critical need for effective communication rehabilitation in individuals who have experienced trauma, such as head injuries or war-related incidents. Tasks include reviewing existing research, selecting appropriate machine learning methods, generating relevant training data, and implementing a software architecture tailored to analyze and correct pronunciation defects in this specific population. The analysis of machine learning methods led to the selection of two experimental models: a Convolutional Neural Network (CNN) utilizing mel-spectrograms for image-based sound representation and a Long Short-Term Memory (LSTM) network combined with mel-frequency cepstral coefficients, aiming to explore the effectiveness of sequential data processing in the context of pronunciation disorder classification in post-traumatic military patients. The results of the two models were compared based on the loss and accuracy functions of the training and validation data, error matrices, and such key metrics as precision, recall, and F1-score. Both models showed promising results in classifying dysarthria stages, but the CNN model performed slightly better in predicting all classes than the LSTM.