CNN-Based System for Voice Disorder Severity Assessment Using Speech Signals

Abstract

Voice disorders in children can affect communication, speech development, and academic performance, making early identification and severity assessment important. This study proposes an Intelligent CNN-Based System for Voice Disorder Severity Assessment Using Speech Signals for children below 12 years of age. The study was conducted to develop an accessible, non-invasive, and cost-effective automated approach for assessing voice-disorder severity using short speech recordings obtained through commonly available mobile devices. Speech samples of 0.2–0.8 seconds were recorded using a mobile phone in a controlled recording room. From each speech signal, a combined set of acoustic features, including Mel-Frequency Cepstral Coefficients (MFCCs), Jitter, Shimmer, and Harmonics-to-Noise Ratio (HNR), was extracted. The extracted features were used to develop four classification models: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), and Convolutional Neural Network (CNN). Experimental evaluation demonstrates that the proposed framework achieves an accuracy of approximately 95%, indicating its ability to distinguish different severity levels from voice recordings. Compared with conventional machine-learning approaches. The models make use of three categories for severity classification of voice disorders namely Low, Moderate, and High. Data augmentation was employed to improve the generalization of the model and tackle the class imbalance. The performance of both the proposed and the comparative model was evaluated using several performance metrics including accuracy, sensitivity, specificity, precision, and F1-score. Results from the experiment illustrate the effectiveness of machine learning and deep learning techniques in identifying severity. The proposed CNN-based system provides a promising approach for rapid and automated preliminary voice-disorder severity assessment in children and may support early screening in resource-limited and non-clinical environments.