Emotion recognition using voice characteristics of speech recordings

摘要

Nowadays, emotion recognition system is one of the most active research topics with a large variety of real-life applications. These applications are diverse, including robotics, education, healthcare, etc. Emotion recognition can be done using different supports such as voice/speech, facial expression, text. These sources can be used separately or together, depending on the cases (e.g., using text and speech to recognize the different emotions expressed by two people during a conversation). This paper deals with the classification of eight different emotions and the classification of one emotion against all the others using speech recordings only. The models have been evaluated using the RAVDESS English database. Supervised machine learning algorithms were used, mainly the cubic SVM algorithm which gives the best results, and global features. An attempt to reduce the number of features using PCA and MRMR algorithm was also made. For the classification of eight emotions, the cubic SVM model keeping the top 200 features using the MRMR algorithm has a weighted average accuracy of 72.1%, achieving the best result.