Recognizing Speech Emotion Based on Acoustic Features Using Machine Learning

Detecting emotion from speech can be helpful to understand the state of individual's mind. Accurately classifying emotion from speech is a very challenging job. In this work, we combined two datasets- Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) and Toronto Emotional Speech Set (TESS) to diversify the speech dataset. The resulting dataset contains 4048 audio files. Seven key emotions of human have been considered for classification including happy, angry, sad, neutral, fearful, disgust, and surprised. 180 speech features have been extracted from audio files utilizing Mel-Frequency Cepstral Coefficient (MFCC), Chroma, and Mel Spectrogram techniques. We applied several traditional classifiers on the combined dataset as well as RAVDESS and TESS datasets separately. Comparative investigation shows that Gradient Boosting outperforms other classifiers on the combined dataset with an accuracy of 84.96%. Also, MLP classifier performs better on all three datasets compared to other classifiers. We believe that this study can contribute to the avenue of human-computer interaction as well as other applications by more precise emotion recognition.