A HYBRID CONVOLUTIONAL NEURAL NETWORK AND SUPPORT VECTOR MACHINE FOR DYSARTHRIA SPEECH CLASSIFICATION

Dysarthria is a neurological disorder that hinders the sufferers to articulate speech properly. These days, Automatic Speech Recognition (ASR) is being researched and developed to help dysarthria sufferers communicate. One of the basic stages of building an ASR is the speech classification and prediction process. In this study, we introduce a CNN-SVM hybrid model to recognize a 10-digit number pronounced by persons with dysarthria. This hybrid model was built to improve the classification ability of a simple CNN architecture in predicting dysarthric speech. CNN is used to capture the unique spatial features from the audio. The features captured by the CNN are then classified by the SVM, as SVM is known for processing data with large features. We also compared our hybrid model with standard CNN. This study succeeded in proving that the hybrid model was better than CNN with softmax layer, with an average increase in accuracy of 7.5%.