SEGMENT-LEVEL MANDARIN CHINESE SPEECH-BASED COGNITIVE IMPAIRMENT DETECTION VIA AN AUTOENCODER WITH CONTRASTIVE LEARNING

Cognitive impairment detection via speech analysis is a promising, non-invasive diagnostic tool. However, clinical implementation is often hindered by the scarcity of labeled data and the high variability across different speech datasets. This paper addresses these challenges by proposing a segment-level representation learning framework designed to extract robust features from speech recordings, even when data is limited.

The researchers converted speech recordings into short segments and transformed them into spectrogram representations. To maximize the utility of limited data, they employed a combination of offline and online data augmentation techniques. The core of the framework utilizes an autoencoder architecture integrated with contrastive learning objectives. This approach forces the model to learn discriminative latent representations, ensuring that the features extracted are highly sensitive to the subtle acoustic markers associated with cognitive decline.

The proposed framework was evaluated across four independent Mandarin Chinese speech datasets. The results demonstrate that the model achieves stable and competitive performance in both binary classification and the more clinically demanding three-class classification tasks. The authors highlight that the segment-level approach provides significant improvements in classification accuracy compared to traditional methods, suggesting that this technique is well-suited for resource-constrained clinical environments where large, high-quality datasets may not be available.

By focusing on segment-level learning and contrastive objectives, this research provides a scalable path toward automated, low-cost screening tools for cognitive impairment. This is particularly relevant for clinical settings where rapid, objective, and non-invasive assessment is needed to support early diagnosis and monitoring.