Individuals with Mild Cognitive Impairment exhibit significant cognitive decline, with a decrease in language expression ability and a higher probability of developing into dementia. Assessing patients’ language abilities, speech expression, and comprehension can aid in the early diagnosis of dementia. In the research, we built a multimodal deep learning model to predict mild cognitive impairment through the symptoms of cognitive decline. The data includes spontaneous speech samples consisting of image description recordings in the TAUKADIAL dataset. We used BERT for audio-to-text conversion and the OpenSMILE library to extract audio features. The features from text and audio were fused, and the feature importance index was calculated as the average of the Support Vector Machine, Logistic Regression, and Random Forest models. Our model achieved better performance on MCI detection in Chinese than the existing methods.