A New Hybrid Feature Extraction and Selection Method based on Deep Learning for Speech Emotion Recognition

In particular, emotion recognition from audio signal is important to enhance human-machine interaction in the current world.They go on to power everything from voice assistants to automated customer support to even healthcare services.We propose a new machine learning model for audio emotion recognition.For feature extraction, a hybrid model is proposed, in which deep features are extracted via a Convolutional Neural Network (CNN) and acoustic features through the Mel-Frequency Cepstral Coefficients (MFCC) algorithm.Afterward, Neighborhood Component Analysis (NCA) algorithm selects the best features concerning the various types of emotions from the combined feature vector and feed them into the Bidrectional LSTM (BiLSTM) for classification.BiLSTM model understand better the temporal relations between the features; therefore, it increases the accuracy of identifying the types of emotions from features events selected.In this study, we evaluated performance of proposed method on three benchmark datasets SAVEE, EmoDB, and RAVDESS.The proposed method effectively extracts and selects discriminative emotional features, enhancing classification accuracy.Experimental results demonstrate the effectiveness of our method, achieving an average accuracy of 88.74% on SAVEE, 95.86% on EmoDB, and 97.02% on RAVDESS.These results indicate the robustness and generalization capability of our model across different speech emotion datasets.