Automatic detection and assessment of Alzheimer Disease using speech and language technologies in low-resource scenarios

In this study, we analyze the use of speech and speaker recognition technologies and natural language processing to detect Alzheimer disease (AD) and estimate mini-mental status evaluation (MMSE) scores. We used speech recordings from Interspeech 2021 ADReSS(o) challenge dataset. Our work focuses on adapting state-of-the-art speaker recognition and language models individually and later collectively to examine their complementary behavior for the tasks. We used speech embedding techniques such as x-vectors and prosody features to characterize the speech signals. We also employed automatic speech recognition (ASR) with interpolated language models to obtain transcriptions used to fine-tune the BERT models that classify and assess the speakers. Our results indicate that the fusion of scores obtained from the multiple acoustic and linguistic models provides the best detection results, suggesting that they contain complementary information. A separate analysis of the models indicates that linguistic models outperform acoustic models in detection and prediction tasks. However, acoustic models can provide better results than linguistic models under certain circumstances due to the errors in ASR transcriptions, which indicates that the performance of linguistic models relies on the performance of ASRs. Our best models provide 84.51% accuracy in automatic detection of AD and 3.85 RMSE in MMSE prediction.