A multi-scale convolutional attention neural network based on residual block downsampling for infant cry classification and detection

The cries of infants contain rich information, indicating hunger, tiredness, discomfort, and other physical discomforts. If we can understand a baby's cry, we can quickly assess their state and take appropriate action. The classification of infant cry sounds plays a crucial role in areas such as household prompts and medical examinations. However, the diversity and complexity of infant cry categories present significant challenges for deep classification design. To improve classification accuracy while conserving computational resources, this study proposes a classification method based on a multi-scale convolutional attention neural network with residual downsampling. The proposed model consists of four parts: a feature extraction module, SE attention module, residual downsampling module based on residual networks, and multi-scale convolution module. Mel spectrogram transformation is applied to the audio to extract feature information, and downsampling is performed to obtain richer feature representations. The combination of multi-scale convolution and attention is used to extract key information from the features. The proposed model achieves an accuracy of 86.29% on publicly available datasets. The model parameters of the proposed algorithm amount to 6.67 million, with a computational cost of 0.66G FLOPs per second. Compared to other existing deep learning algorithms, the proposed model achieves the highest accuracy while considering both the parameter quantity and computational cost simultaneously.