Analysis of Emotions from Speech using Hybrid Deep Learning Network Models

In today’s advanced digital world, Emotion plays a important role for communication to understand better each other in virtual environment and for Human to Machine Communication. Research has revealed the powerful role that emotion plays in shaping human social interaction. This has opened up a new research field called Automatic Emotion Recognition, having basic goal to understand and retrieve desired emotions. In the past several biometrics have been explored to recognize the emotional states such as facial expressions, speech, and physiological signals, etc. However, several inherent advantages make speech signals are good source for affective computing. The primary objective of this study is to assess various emotions from speech using Deep learning algorithms (CNN and LSTM). Traditional/Machine Learning approaches rely on manual feature extraction before classifying the emotional state, whereas Deep Learning networks widely used for emotional classification due to their advanced feature extraction mechanisms. In this implementation, we will use the most advanced hybrid deep learning models 1-D CNN+LSTM and 2-D CNN+LSTM, to extract features and classify seven distinct emotions based on their input data. Using the Emo-DB standard dataset, we trained and test the proposed network model and achieved a maximum accuracy of 91.51% and 95.52% for I-C CNN+LSTM and 2-D CNN+LSTM respectively.