- Linguistic and acoustic features are extracted from speech and language using both hand-crafted and automatic methods. Hand-crafted features include pauses and parts of speech while automatic features use models like BERT. - Models are developed for tasks like detecting cognitive impairment across languages and in low-resource settings. Semi-supervised models are created to handle limited labeled data. Models also aim to remove bias from non-clinical factors like age. - Quality assurance examines the effects of errors like those from automatic speech recognition systems. The impact of heterogeneous, multi-task datasets is also analyzed to improve model performance for detection tasks.