Dysarthric Speech Recognition using Multiple Voice Features and Hybrid Deep Learning Classifier

Dysarthric speech recognition (DSR) is difficult due to the large variance in voice quality and limited comprehensibility. To increase the DSR’s accuracy, dysarthric speech must have a better feature representation. This study introduces the new hybrid DLGNet, which combines the Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Deep Convolution Neural Network (DCNN). The DCNN provides high-level hierarchical characteristics to describe the dysarthria-related alterations in voice. The LSTM provides the voice’s temporal representation and long-term dependence. The GRU provides a high-level, long-term reliance and correlation between the local and global aspects of the dysarthric voice. For DSR, the score level fusion is used. Using the publicly available UASpeech dataset, the DSR system’s performance is assessed in terms of F1 score, accuracy, recall, and precision. The DLGNet offers better feature depiction and performs better for low intelligibility Dysarthric speech. The proposed method provides recall of 0.98, precision of 0.97, F1-score of 0.98, and accuracy of 97.25 %, which has shown a noteworthy boost over the traditional state of arts.