Early detection of mild cognitive impairment (MCI) and Alzheimer’s disease (AD) is crucial for timely intervention, yet existing cognitive screening tools are often time–consuming, resource–intensive, and culturally biased. We introduced a Mandarin spontaneous speech dataset collected from elderly participants in inland China, accompanied by demographic, affective, and cognitive assessments. We further designed a culturally adapted picture description stimulus (Chinese Kitchen) addressing limitations of the widely used Cookie Theft picture. Leveraging both handcrafted paralinguistic features (eGeMAPS) and self-supervised speech representations (HuBERT and wav2vec 2.0 XLSR-53), we benchmarked several contemporary pipelines for speech-based cognitive screening on our dataset, including MLP, Bi-LSTM, and a CNN-Transformer backend. Experiments with 10-fold cross-validation demonstrated that self-supervised features substantially outperformed engineered features, proving the effectiveness of our culturally adapted stimulus at the same time. Future work may extend to dialect robustness and multimodal fusion research (speech-gesture fusion) via synchronized high-quality audio-video.