An Accurate Speech Emotion Analysis using Gradient Boosting and BiLSTM Techniques

In the field of speech emotion recognition, machine learning (ML) algorithms face a challenging problem in recognizing emotional states. Speech emotion recognition (SER) is crucial in many real-time solutions, including human-robot interaction, assessing human behavior, and analyzing the emotional state of speakers. Due to its intricacy, SER is a task that is still difficult to complete but one that is becoming more and more crucial in emotional computing. Although there are many datasets available for SER, we are selecting RADVESS since we believe it to be the most ap-propriate. For many applications, such as a speech-to-speech translation system, feature selection can increase the accuracy of speech emotion recognition. Although tens of thousands of features can be retrieved from speech signals, it can be challenging to determine which elements are most closely associated to speaker emotional state. Most characteristics that are connected to emotional states are yet to be discovered. Mel Spectrogram and MFCC feature selection are the major topics of this research. Then passing these combined features through a modified CNN. In this paper we have used various neural network techniques (NN) for SER. The models we used include CNN, MLP, CNN-Attention-BiLSTM, Gradient Boosting. CNN-Attention- BiLSTM gives the best results among these models with accuracy of about 86 percent, precision of 0.834, recall of 0.802, F1-Score of 0.802.