Cross Lingual Speech Representation for Infant Cry Classification

Infant Cry Classification is a critical area of research for medical diagnosis, early detection, and timely intervention. Classification of normal vs. pathological infant cries can be of potential use to infer the infant’s health conditions. In this study, we offer a unique method that employs deep learning models, such as wav2vec, HuBERT, and XLS-R, as well as Whisper feature-based classification. We used Bi-directional Long-Short Term Memory (BiLSTM) as a deep learning classifier in this investigation. We compared XLS-R 0.3B’s extracted features to those of other frequently used pre-trained models. Our proposed approach claims 27.6 % increase over wave2vec 2.0 features on standard Baby Chillanto dataset. In addition, a study of latency period is conducted to assess the practical impact of the proposed XLS-R 0.3B characteristics. Finally, the noise robustness of the suggested feature vector has been investigated in this study.