Feature selection using CNN for elderly speech recognition

Functional decline is a severe syndrome that affects older persons, and its early risk identification can postpone or prevent its consequences. Speech recognition is employed in this study to assist with the self-evaluation of the geriatric tests used for functional decline risk assessment. We use Convolutional Neural Network CNN model as a recognizer for word recognition. We apply feature selection and sub-feature selection techniques on extracted features to select the most convenient feature for the CNN and reduce the input feature matrix. Our findings demonstrate that the combination of MFCC, RASTA-PLPc, and PLPs gave better WRR=96.50% than each one alone and excluded other features. It also showed that the order of the features affects the model’s performance. Similarly, the sub-feature selection is feasible and increases the WRR to 98.95 %.