Transformer-Based Infant Cry Classification for Intelligent Neonatal Monitoring Systems

Resumen

Infant cry analysis is often put in the context of traditional audio classification. However, the reductionist formulation often hides structural issues that abound in the real world such as a high level of class imbalance, overlapping features of acoustics and a low level of ecological validity of the datasets involved. In neonatal care facilities that are operating (e.g. smart incubators or intelligent monitoring systems) cry signals do not adhere to a controlled lab setting. To counter such deficits, the current paper proposes a transformer-based sequence of modeling of multi-class infant cry recognition with a set of eight behavioral states. The methodology clearly presents the imbalance of datasets through a synthesis of data augmentation based on controlled data, class-weighted optimisation, and termination, thus stabilising the training of models instead of artificially inflating the performance measures.The system uses a hybrid acoustic model which combines timefrequency structures with complementary spectral and prosodic features. This collection of composite features enables the transformer encoder to obtain short-term acoustics as well as long-term dependencies that characterize infant crying behavior. The imbalance-conscious metrics (Macro-F1 score and balanced accuracy) are made the priority of the model evaluation to avoid interpretive bias introduced by the dominant classes. The experimental findings indicate that the suggested structure can achieve the overall accuracy of 96.27% and maintain the steady performance across the minority cry types. Despite these positive observations, the study presents limitations that are related to the size of datasets and possible subject-level overlap of data. Under these limitations, the results suggest that transformer based cry state inference can serve a reliable decision support component of intelligent infant monitoring system, but it cannot be considered an autonomous diagnostic resource.