Indonesian speech emotion recognition: feature extraction and neural network approaches

Abstract

This study explored the challenges of emotion recognition in Indonesian speech using deep learning techniques, addressing the complex nuances of emotional expression in spoken language that posed significant difficulties for automatic recognition systems. The research focused on the application of feature extraction methods and the implementation of convolutional neural networks (CNN) and a hybrid convolutional neural networks-long short-term memory (CNN-LSTM) model to identify emotional states from speech data. By analyzing key features of speech signals, including mel frequency cepstral coefficient (MFCC), zero crossing rate (ZCR), root mean square energy (RMSE), pitch, and spectral centroid, the study evaluated the models’ ability to capture both spatial and temporal patterns in the data. Testing was conducted using an Indonesian dataset comprising 200 samples. The CNN model, utilizing four features (MFCC, ZCR, RMSE, and pitch), and the CNN-LSTM model, which used three features (MFCC, ZCR, and RMSE), both achieved an emotion classification accuracy of approximately 88%. The result showed that the CNN-LSTM model achieved comparable performance with a simpler feature set compared to the CNN model. This highlighted the significance of choosing the appropriate techniques in feature extraction and classification to enhance the accuracy of identifying emotions from speech data while also managing computational complexity.