Audio Emotion Recognition using machine learning algorithms
Speech is an important part of human to human interaction and can be an even more important part of human to computer interaction where speech is the only source of communication. Detection of emotion from speech is an emerging area of research in the field of Human Computer Interaction as it can improve it in a lot of fields like customer service centres, entertainment or utility electronics which work with speech detection, text to speech systems etc. This project develops an Audio Emotion Recognition system which detects emotions using audio signal features like Mel-Spectogram, Mel-Frequency Cepstral Coefficients(MFCC) and Chroma.
This project presents a comparision of the performance of three such systems namely the Multi Layer Perceptron(MLP), Convolutional Neural Network(CNN) and CNN- Long-Short Term Memory(LSTM). Experiments are conducted using a single dataset and also with combination of different datasets(RAVDESS, SAVEE, CREMA, TESS). Results indicate that CNN-LSTM performed better than the other models.