Research on Pathological Voice Recognition using Multi-Scale Convolutional Neural Network

The performance of existing neural network models for pathological voice recognition no longer meets the requirements. They struggle to address the diversity in pathological voice features. There is an urgent need to find a neural network model with better performance and stronger robustness. The multi-scale convolutional neural network (MSCNN) model possesses the powerful capability to explore multi-scale convolutional blocks to extract multi-scale representations for detection, and it has achieved success in various recognition applications. In this paper, we propose the MSCNN model for pathological voice recognition. It utilizes multi-scale convolutional blocks for feature extraction from spectrograms of voice signals and employs fully connected layers as classifiers. We compare the MSCNN model with a one- dimensional convolutional neural network (lDCNN), a long short-term memory network (LSTM), and a two-dimensional convolutional neural network (2DCNN) as baseline methods. The IDCNN and LSTM methods process voice signals that have been framed and windowed as input, while the 2DCNN and MSCNN methods take spectrograms of voice signals as input. Experiments are conducted on three different databases: MEEI, SVD, and HUPA. We evaluate the performance of the models using five metrics: accuracy, precision, recall, F1 score, and Matthews correlation coefficient (MCC). The experimental results demonstrate that MSCNN exhibits excellent performance in pathological voice recognition. In the MEEI, SVD, and HUP A databases, it achieves binary classification accuracies of 95.087%, 69.809%, and 75.455%, respectively, surpassing the other three methods. Moreover, it achieves favorable results in precision, recall, F1 score, and Matthews correlation coefficient.