Predicting Hypernasality Using Spectrogram Via Deep Convolutional Neural Network (DCNN)

Hypernasality, marked by excessive nasal resonance during a speech, impairs communication clarity, particularly for sounds requiring soft palate closure, such as the “f” in ‘football’. Common causes include cleft lip and palate, velopharyngeal insufficiency, neurological disorders, adenoid issues, anatomical defects, and nerve system impairments. Detection and management in clinical settings are challenging due to the limited availability of speech-language pathologists (SLPs), subjective evaluation methods, and inefficient workflows. This study uses a speech feature with a deep learning model to predict the hypernasality in individuals. The objective of the study is to reduce reliance on SLPs, ensure consistent and objective evaluations, and enhance clinical efficiency. The study achieved a $\mathbf{9 2 . 3 1 \%}$ accuracy in detecting hypernasality using the spectrogram features extraction method via Deep Convolutional Neural Network (DCNN). The clinical can apply this study by developing a user-friendly GUI and integrating it with a secure database. Using suitable tools for deployment, this study can be effectively applied in clinical settings. The contribution of this study is that it addresses the scarcity of SLPs and enhances clinical evaluation practices, ultimately improving communication outcomes for individuals with hypernasality.