Feature Extraction Technique Based on Conv1D and Conv2D Network for Thai Speech Emotion Recognition

Speech Emotion Recognition is one of the challenges in Natural Language Processing (NLP) area. There are many factors used to identify emotions in speech, such as pitch, intensity, frequency, duration, and speakers' nationality. This paper implements a speech emotion recognition model specifically for Thai language by classifying it into 5 emotions: Angry, Frustrated, Neutral, Sad, and Happy. This research uses a dataset from VISTEC-depa AI Research Institute of Thailand. There are 21,562 sounds (scripts) divided into 70% of training data and 30% of test data. We use the Mel spectrogram and Mel-frequency Cepstral Coefficients (MFCC) technique for feature extraction and 1D Convolutional Neural Network (Conv1D) all together with 2D Convolutional Neural Network (Conv2D), to classify emotions. With respect to the result, MFCC with Conv2D provides the highest accuracy at 80.59%, and is higher than the baseline study, which is of 71.35%.