Deep Learning based Dysarthric Speech Recognition using Spectral, Temporal, and Phonetic Acoustic Features

Dysarthric speech recognition (DSR) is very challenging due to the vast variability and low intelligibility of the voice, which is associated with distinct neurological disorders. Traditional speech recognition systems fail to identify speech with dysarthria due to the wide variation in prosody and intonation. Various systems have been developed for detecting DSR, but they are limited by lower generalization, suboptimal feature representation, and lower recognition rates. This paper presents the DSR using a Deep Convolutional Neural Network and Long Short-Term Memory (DCNN-LSTM), which provides better feature depiction, superior temporal correlation, and long-term connectivity in the features. The DCNN-LSTM helps boost the discriminative properties of the voice to characterize the prosodic, intonational, spectral, and temporal variations associated with dysarthria. The effectiveness of the DSR scheme is evaluated on the UASpeech dataset for digit recognition, achieving 93.88% accuracy.