Machine Learning Based Emotion Detection

Abstract

The detection of emotion in speech has become relevant as social media becomes more popular and audio-based attacks become more frequent. This project aims to identify five emotions, including happiness, sadness, threat, panic, and neutrality, in the Bangla speech. The 1,106 five second audio samples dataset was formed by using different sources and was also extended to 6,636 samples by inserting various types of noise. The MFCC, Melspectrogram, Spectral Centroid, and LPC features had been extracted and used by both machine learning and deep learning models. Random Forest yielded the best results among conventional algorithms, with a 1D Convolutional Neural Network (CNN) performing better with the larger dataset, achieving 83 percent test accuracy and large ROC values. Gender-based testing was more accurate in female voices (88.07%) and male voices (81.2%). Further overfitting was suppressed and accuracy stabilized by the use of higher-order statistics, including the Teager Energy Operator. Finally, the project was able to construct a Bangla speech sentiment database and create an effective CNN-based classifier. The findings prove to have potential uses in security, psychological evaluation, and social media surveillance.