Thai Speech Emotion Recognition Using Artificial Neural Networks

The research aimed (1) to develop and evaluate an emotion recognition model for Thai speech using artificial neural networks, (2) to enable accurate classification of human emotions, and (3) to bridge communication gaps between computers and users. A dataset from AIResearch.in.th consisting of 27,854 Thai-language sentences categorized into angry, sad, happy, frustrated, and neutral emotions. The Mel Frequency Cepstral Coefficients (MFCC) employed for the speech feature extraction. Data was pre-processed by augmentation techniques, including time stretching, pitch shifting, and noise injection. The pre-processed data trained for artificial neural network models, including a 1-dimensional Convolutional Neural Network (1D CNN), Long Short-Term Memory (LSTM), and a hybrid model (1D CNN and LSTM). Results showed that the hybrid model (1D CNN & LSTM) achieved the highest accuracy of 80.36%, followed by the 1D CNN model (77.52%) and the LSTM model (67.86%).