Learning Salient Features for Speech Emotion Recognition using Attention-based Residual Bidirectional LSTM with Federated Learning

Abstract

In human communication, expressing human emotion plays an essential role in transferring information to the other individual. The expression forms of human communication are very rich, in different patterns like facial expressions, voice tone, eye contact, body language and laughter. Moreover, the languages used by individuals in the entire world are different, yet without understanding the communication language, the individuals can realize the message that other individuals require to convey as emotional expressions. Among all the human emotional expressions, the conveyed expressions through human voice are experimented with in different studies. However, identifying the emotional situation of the presenter is complicated for the machine learning techniques, which gained more attraction in Speech Emotion Recognition (SER). Here, the SER framework plays an efficient role in different real-world applications in observing human actions, virtual reality, emergency centers and robot communication to observe the emotional phase of the speakers. Thus, an innovative federated learning-based SER is presented in this paper. The essential speech signals are initially collected from the standard database. After that, the collected signal is fed to the feature extraction stage. Here, the deep features in the input signal are obtained by the Variational Autoencoder (VAE). Further, the deep features obtained from the VAE are passed to the "Attention-based Residual Bidirectional Long Short-Term Memory (A-RBiLSTM)" to identify the emotions in the speech signal. At last, different experimental analyses are executed to observe the effectualness of the implemented framework.