Performance Investigation of Customized MFCC Feature Extraction in Recognizing Indonesian Conversational Emotion

Modifying the MFCC extraction process is an opportunity to improve Indonesian SER performance. This work confirms the performance of MFCC features obtained by customized MFCC extraction. This paper utilizes customized MFCC features with single-layer LSTM as a classifier model. MFCC was customized by providing values of pre-emphasis of 0.97, a frame length of 0.025 seconds with an overlap of 0.01 seconds, FFT points of 512, a mel-filter of 40, and 13 MFCCs. The result shows better performance in single-layer LSTM compared with SVM using PCA, which results in a training accuracy of 0.789 and a testing accuracy of 0.787. K-fold cross-validation was confirmed with an average accuracy of 0.816 using k value 5 and 0.879 with k value 10.