Classification of Infant Cry Based on Hybrid Audio Features and ResLSTM.

ABSTRACT

Crying is one of the primary means by which infants communicate with their environment in the early stages of life. These cries can be triggered by physiological factors such as hunger or sleepiness, or by pathological factors such as illness or discomfort. Therefore, analyzing infant cries can assist inexperienced parents in better caring for their babies. Most studies have predominantly utilized a single-speech feature, such as Mel Frequency Cepstral Coefficients (MFCC), for classifying infant cries, while other speech features, such as Mel Spectrogram and Tonnetz, are often overlooked. In this study, we manually designed a hybrid feature set, MMT (including MFCC, Mel Spectrogram, and Tonnetz), and explored its application in infant cry classification. Additionally, we proposed a convolutional neural network based on residual connections and long short-term memory (LSTM) networks, termed ResLSTM. We compared the performance of different deep learning models using the hybrid feature set MMT and the single MFCC feature. This study utilized the Baby Crying, Dunstan Baby Language, and Donate a Cry datasets. The results indicate that the hybrid feature set MMT outperforms the single MFCC feature. The MMT combined with the ResLSTM method achieved the best performance, obtaining accuracy rates of 94.15%, 92.92%, and 95.98% on the three datasets, respectively.