Infant Crying Recognition Method of Limited Data Based on Self-supervised Learning

Infant cry recognition and interpretation are crucial for understanding and addressing infant needs. However, current supervised learning methods suffer from limited annotated infant cry data, leading to performance degradation in classification. In this paper, we propose a self-supervised learning approach for infant cry analysis. The proposed method involves masking the speech input within the latent space and solving a contrastive task defined over quantized latent representations which are jointly learned. By leveraging a substantial volume of unlabeled data, the model discerns crucial audio contrast features. Subsequently, fine-tuning is performed by using a small set of labeled data. The unlabeled data comes from the 960h audio data of the Librispeech dataset. The labeled infant crying data consists of 926 audio tracks (hungry (565), discomfortable (195), angry (82), disturbed (84)), sourced from major hospitals and postpartum clubs, hich achieves accuracies of 98.67%, 95.60%, and 95.18% in binary, ternary, and quaternary classifications, respectively. Compared to traditional supervised learning models (CNN, TDNNS, Res2Net), there is a significant improvement in accuracy.