Vocal Performance Evaluation System using Hybrid Convolutional Neural Network and Long Short-Term Memory

The integration of deep learning with voice analysis has advanced vocal performance evaluation, especially in diagnosing voice disorders. Accurate assessment of vocal health is vital for early detection and effective treatment. This research presents a Vocal Performance Evaluation System using the VOICED dataset, which includes diverse speech recordings with disorder annotations. While automatic voice assessment has progressed, existing methods often struggle to capture both spectral and temporal features, affecting classification accuracy. To overcome this, a fusion-based neural framework is introduced, utilizing Convolutional Neural Networks (CNNs) for spectral feature extraction and Long Short-Term Memory (LSTM) models for temporal modeling. Mel-Frequency Cepstral Coefficients (MFCCs) are employed as key features to capture the intricate spectral characteristics of speech signals. The proposed model demonstrates superior performance attaining a classification accuracy of 98.7%. Evaluation metrics, including Precision (98.4%), Recall (97.9%), Specificity (99.1%), and F1-Score (98.2%), confirm the model’s effectiveness. Experimental results highlight the framework’s strength in voice disorder detection, offering notable improvement over existing approaches. These findings underscore the value of AI-driven voice assessment systems in clinical and real-world use, contributing to early diagnosis and better treatment outcomes.