Feature Extraction And Machine Learning Classifiers For Elderly Speech Recognition In Comprehensive Geriatric Assessment Cga Questionnaires

Functional decline is a serious syndrome that attacks older adults and can be delayed or prevented by its early risk detection. In this study, speech recognition is used to help in the self-assessment of the geriatric tests used for this purpose. We investigate the effect of the speaker mode, the frame length, the speaker gender, and the vocabulary size in terms of the word recognition rate (WRR) for three different classifiers: RF, KNN, and SVM. The results of our data show that all three classifiers are proper for speech recognition. However, for speaker-dependent mode, the RF gives the maximal performance when separating genders (for females 98.6% and for males 98.1%) and using a 25 ms frame-length with a 10 ms overlap. On the other hand, for the speaker-independent mode, the RF gave also the best performance for 20ms with 10ms overlap, for digits alone WRR=98.1% and binary classification alone WRR=98%.