Automatic Diagnosis of Hyperkinetic Dysphonia from Speech Recordings Based on Deep Learning Approaches

The automatic diagnosis of hyperkinetic dysphonia using deep learning techniques represents a significant advancement in voice disorder assessment. This approach leverages the analysis of spectrograms, Mel-spectrograms, and Mel-Frequency Cepstral Coefficients (MFCCs) extracted from speech recordings to enable precise diagnosis. Hyperkinetic dysphonia is a voice disorder characterized by excessive vocal fold tension, traditionally diagnosed through subjective assessments and invasive procedures. These conventional methods are often time-consuming, uncomfortable for patients, and prone to inconsistencies, which can lead to potential delays in treatment. In this study, various deep learning models, including multilayer perceptron (MLP), convolutional neural networks (CNN), and long short-term memory (LSTM), a specialized type of recurrent neural network (RNN) are employed to offer a non-invasive, efficient, and highly accurate diagnostic alternative. The analysis is conducted on healthy and hyperkinetic voice samples from the VOICED dataset. Voice signals are processed and segmented into five and ten segments, followed by the extraction of spectrograms, Mel-spectrograms, and MFCCs for each segment. Each model was trained and evaluated using the bootstrapping technique for robustness. The classification results demonstrate that the accuracy of all models ranges from approximately 67.1% to 99.8%. The regularized MLP model, which utilized Mel-spectrogram features extracted from 10 audio segments, exhibits superior performance across six metrics: accuracy (0.9984), precision (0.9985), F1 score (0.9984), recall (0.9984), specificity (0.9980), and Area Under the Curve (AUC) (0.9999). It outperforms existing methods and other deep learning approaches. These findings emphasize the potential of integrating deep learning models into clinical practice for diagnosing hyperkinetic dysphonia. The proposed approach provides a non-invasive, reliable, and highly accurate diagnostic tool, reducing the workload of healthcare professionals while enhancing patient care outcomes.