Analysis of Multiple Types of Baby Cries Based on LSTM

The literature currently provides limited analyses of multiple emotions conveyed by infant cries and associated training sets. To address this gap, we collected six common infant cry speech types, which include awake, needing a diaper change, needing to be held, hungry, sleepy, and uncomfortable, from online sources and constructed them into a training set. LSTM (Long Short-Term Memory Neural Network) was utilized for the analysis of infant cry speech in this study. Voice features based on spectral and MFCC features were extracted and inputted into the LSTM network, supplemented by three fully connected layers to enhance generalization ability and robustness. Our results demonstrate that the LSTM model achieved a maximum accuracy of 92.39% in identifying multiple types of infant cries while plateauing after 75 rounds, indicating both efficiency and good performance. Additionally, the confidence probability of the validated speech obtained by the model was 92.9%, demonstrating good generalization ability and robustness based on actual speech data. Therefore, our developed LSTM model has good potential in classifying multiple types of emotions conveyed by infant cries with a high degree of accuracy and robustness.