Utilizing CNN Architectures for Non-invasive Diagnosis of Speech Disorders

This study explored the potential of convolutional neural networks (CNNs), like VGG and ResNet, in diagnosing diseases affecting the voice and speech apparatus through non-invasive analysis of vowel sound recordings. Utilizing the Saarbruecken Voice Database, voice recordings were transformed into spectrograms for model training, focusing on the vowels /a/, /u/, and /i/. The research employed Explainable Artificial Intelligence (XAI) techniques to highlight critical features for disease detection within these spectrograms, aiming to provide medical professionals with deeper insights into disease manifestations in voice patterns. Based on the F1-score, performance evaluation revealed that ResNet18 models achieved a score of 71.51 ± 2.49, outperforming VGG19 models, which scored 68.51 ± 1.78. The study’s findings suggested that vowel selection and data augmentation strategies did not significantly enhance model performance. Additionally, it highlighted the inefficacy of using the ‘U’ and ‘I’ vowels as standalone indicators for disease diagnosis, recommending the aggregation of multichannel data to improve feature extraction by the models. (All results are fully reproducible, the source code is available at https://github.com/Tesla2000/DepCoS2024 ).