Laryngeal Disease Detection Using Machine Learning

Voice disorders are frequently underdiagnosed due to the subjective and resource-intensive nature of traditional clinical assessments. This work proposes a non-invasive, machine learning based system for early detection and classification of voice pathologies, which is particularly beneficial in settings where access to specialized clinical expertise is limited and vocal health is critical for daily communication and professional use. Using the Saarbrücken Voice Database (SVD), a two-step classification framework is designed hat first performs healthy versus pathological voice detection and then conducts multi-class classification across six clinically significant pathology categories. Audio preprocessing involves FFmpeg-based conversion and extraction of MFCCs, spectral features, jitter, shimmer, pitch, and formants. The evaluation covered several architectures, spanning classic models like Random Forest and LightGBM alongside neural network frameworks such as BiLSTM, 1D and 2D CNNs, LSTM, and hybrid CNN+LSTM setups. Results demonstrate that LightGBM achieves the best performance in binary classification, while the fine-tuned CNN+LSTM hybrid model provides superior multi-class accuracy. This pipeline lays the foundation for a realtime AI-assisted diagnostic tool for scalable clinical deployment.