Enhancing infant cry recognition using lightweight CNN with hybrid feature augmentation

Abstract: Automated infant cry classification remains challenging in real-world scenarios owing to the limited annotated data and the computational burden of deep learning models. To address these issues, this study proposes a lightweight convolutional neural network (CNNL) combined with a hybrid feature augmentation (HFA) strategy for robust infant cry recognition. The proposed CNNL contains only 6.3M parameters, substantially fewer than the 20.3M parameters of ResNet32, while achieving superior classification performance. Meanwhile, HFA en hances training diversity and model generalization through time shifting, speed variation, pitch transformation, and noise insertion. Experimental results on two benchmark datasets show that the proposed framework out performs several representative baselines, including ResNet32, ResNet18, and MobileNetV2. With HFA, the proposed method achieves classification accuracies of 97.14% on the Babycry dataset and 95.00% on the Donateacry dataset. These results confirm the effectiveness of the proposed method as a compact and high- performing solution for infant cry classification.

Keywords: Infant cry, Lightweight network, Feature augmentation, Feature fusion, Log-Mel spectrogram

Recommended citation: Zhang M, Lu J, Cheng L, et al. Enhancing infant cry recognition using lightweight CNN with hybrid feature augmentation[J]. Biomedical Signal Processing and Control, 2026: 110367.
Download Paper