Towards objective and interpretable speech disorder assessment: a comparative analysis of CNN and transformer-based models

Some research has been focused on using these models to automatically assess Head and Neck Cancers (HNC) significantly impact patients' the speech severity level [13, 14, 15]. Other studies analysed ability to speak, affecting their quality of life. Commonly how well diseases can be predicted by these models. For instance, used metrics for assessing pathological speech are subjective, A. Favaro et al. [16] compared interpretable speech prompting the need for automated and unbiased evaluation features to embeddings produced by SSL models on predicting methods. This study proposes a self-supervised Wav2Vec2-the presence of Parkinson's disease. They showed that based model for phone classification with HNC patients, to enhance using embeddings provides better detection accuracies at the accuracy and improve the discrimination of phonetic features cost of losing the insight into speech and language deterioration for subsequent interpretability purpose. The impact of given by interpretable features. While being able to detect pre-training datasets, model size, and fine-tuning datasets and a disease and assess its severity is important, we believe it parameters are explored. Evaluation on diverse corpora reveals is as important to interpret the output of these models, in order the effectiveness of the Wav2Vec2 architecture, outperforming to enhance trust that clinicians can have in these systems.