Abstract:
This study examines fundamental autonomous speech parameters and incorporates text analysis to detect dementia from voices, while also assisting dementia patients by providing sentence auto completion support to address their communication challenges. To accomplish these 740 voice recordings (370 dementia and 370 non- dementia) were collected. Features such as MFCCs and RMS were extracted from the audio and text data. Various machine learning models, including Random Forest (RF), Logistic Regression (LR) and Gradient Boosting (XGBoost), alongside a deep learning Long Short-Term Memory (LSTM) were trained. Among these, the LSTM model achieved the highest accuracy of 92.93%. The recorded voices were transcribed into text using Whisper model, and TF-IDF trigram features were extracted for detection. The models were implemented for text classification, with LR and LSTM achieving the best accuracies of 72.43% and 72.78% respectively. For sentence auto completion, a Bi-directional LSTM (Bi-LSTM) model with N-gram sequences was implemented and achieved 20.8% accuracy. This research highlights the integration of speech and text-based methods to analyze and detect dementia and assist dementia patients through sentence auto completion.
