Emotion Classification of Infants using Audio Signal Analysis

Abstract

The purpose of this research is to dive deep into infant emotion classification techniques on audio signals to classify infants’ emotions expressed in sound into comfortable or uncomfortable emotions. Audio data to be used for this research has been collected and annotated manually. Most of the data has been collected from friends and family, while some publicly available data also have been used. Audio augmentation techniques such as time stretching, and noise addition have been explored and used appropriately to enrich the dataset. To clean up audio data, audio pre-processing techniques has been applied on the dataset. After extensive literature review, log mel spectrogram feature has been used as the feature matrix for the classifier. This feature is very popular for emotion / sentiment classification related problems as this feature is a close representation of how humans perceive sound. A CNN deep learning model with multiple pooling and dense layers have been used to classify the feature matrix extracted from the audio files. Hyper parameter tuning techniques have been used to maximize the accuracy of the model. The classifier implemented demonstrated an accuracy exceeding 80%, which exceeds most of the similar research accuracies.