Towards Non-Invasive Speech Evaluation: AI-Based Detection of Cleft Lip from Audio Recordings

Cleft lip and/or palate is among the most prevalent congenital malformations globally, significantly impacting speech production and facial morphology. Current clinical assessments of CLP-related speech impairments often involve invasive and subjective procedures, limiting accessibility and consistency. This study introduces a non-invasive, AI-driven approach for automatically classifying speech from individuals with and without CLP, leveraging acoustic embeddings extracted via the Wav2Vec 2.0 model. Two supervised machine learning algorithms—Support Vector Machines (SVM) and neural networks—were trained and rigorously optimized. Experimental results indicate that the neural network outperformed the SVM, achieving superior F1-scores and overall classification performance. The speech corpus, comprising 16 phonetically rich sentences per participant, provided sufficient variability for robust group differentiation. These findings offer a promising foundation for the development of objective, scalable, and accessible tools to support clinical evaluation and early diagnosis of speech anomalies associated with CLP.