Acoustic Analysis for Early Detection of Vocal Disorders

Vocal disorders often manifest through subtle acoustic changes such as hoarseness, pitch instability, and vocal fatigue, which are difficult to detect during early stages using conventional clinical methods. Delayed identification can lead to prolonged treatment and increased health complications. To address this challenge, this paper presents an automated, non-invasive system for the early detection of vocal disorders using deep learning techniques. The proposed approach analyzes recorded voice signals by applying signal preprocessing methods followed by feature extraction using Mel spectrograms and acoustic parameters such as MFCCs, pitch, jitter, and shimmer. A hybrid CNN–LSTM architecture is employed to effectively capture both spatial frequency patterns and temporal dependencies present in speech signals. The convolutional neural network extracts discriminative spectral features from spectrogram representations, while the long short-term memory network models time-dependent vocal characteristics. The system classifies voice samples into normal and disordered categories using supervised learning. Experimental results demonstrate reliable classification performance, indicating the effectiveness of the proposed model as a preliminary screening tool. Although not intended to replace professional medical diagnosis, the system provides a cost-effective and accessible solution to support early identification and timely referral for clinical evaluation.