Comparative analysis of deep learning models for dysarthric speech detection

Dysarthria is a speech communication disorder that is associated with neurological impairments. To detect this disorder from speech, we present an experimental comparison of deep models developed based on frequency domain features. A comparative analysis of deep models is performed in the detection of dysarthria using scalogram of dysarthric speech. Also, it can assist physicians, specialists, and doctors based on the results of its detection. Since dysarthric speech signals have segments of breathy and semi-whispery, experiments are performed only on the frequency-domain representation of speech signals. Time-domain speech signal is transformed into a 2-D scalogram image through wavelet transformation. Then, the scalogram images are applied to pre-trained convolutional neural networks. The layers of pre-trained networks are tuned for our scalogram images through transfer learning. The proposed method of applying the scalogram images as input to pre-trained CNNs is evaluated on the TORGO database and the classification performance of these networks is compared. In this work, AlexNet, GoogLeNet, ResNet 50 and two pre-trained sound CNNs, namely VGGish and YAMNET are considered deep models of pre-trained convolutional neural networks. The proposed method of using pre-trained and transfer learned CNN with scalogram image feature achieved better accuracy when compared to other machine learning models in the dysarthria detection system.