Deep Learning Based Speech Recognition for Hyperkinetic Dysarthria Disorder

Speech recognition is a technology that aims to transform human speech into text and has applications in a variety of fields, including information technology, healthcare, automobiles, and others. This research explores the advantages of employing deep learning methods to provide individualized assistance technology solutions for people with dysarthria. We proposed a Convolutional Temporal Bidirectional Network (CTBNet) model for translating Russian Hyperkinetic Dysarthria Disorder (RHDD) speech into text. CTBNet encodes input audio features using 1D convolution layers, and BidirectionalGRU (Bi-GRU) layers. The encoder effectively captures both short-term and long-term spatial temporal information. More importantly, CTBNet includes with an attention-CTC decoder, that is responsible for producing output texts. We utilized a dataset of 2797 recordings of RHDD speech, with a cumulative recorded duration of 2 hours and 33 minutes, for training purposes. The CTBNet model has achieved a 15% Character Error Rate (CER) and a 59% Word Error Rate (WER) on the dataset proposed. Our experimental findings show superior performance compared to state-of-the-art models, namely, 3xCNN, and 3xLSTM, when evaluated on the same dataset.