SPEECH EMOTION RECOGNITION USING SPEECH PROCESSING: A HYBRID CNN–BILSTM DEEP LEARNING APPROACH WITH FEATURE FUSION

Abstract

Speech Emotion Recognition (SER) is a cutting-edge field under the affective computing framework. Its core function is to identify five basic human emotions as happiness, sadness, anger, fear, and surprise from speech input, to support the optimization of human-computer interaction experiences. Early traditional machine learning methods adopted in this field relied on manually designed acoustic features such as MFCC, fundamental frequency, energy, and spectrum. These methods suffered from insufficient generalization in complex cross-speaker, cross-language, and noisy scenarios. In recent years, deep learning methods including CNN, RNN, LSTM, and their hybrid architectures can automatically extract discriminative speech representations, greatly improving the overall performance of SER. This study proposes a CNN-BiLSTM hybrid architecture with fused features: the CNN module extracts spatial features from spectrograms, while the BiLSTM module captures the temporal relationships of speech sequences. Five universal benchmark datasets like RAVDESS, TESS, CREMAD, SAVEE, and EMO-DB are selected to conduct model evaluation. Three data augmentation techniques, namely noise injection, fundamental frequency shifting, and time stretching, are adopted to enhance model robustness. The experiments achieved an accuracy rate of 93%-98%, outperforming all traditional and mainstream deep learning benchmark models. Ablation experiments verified the necessity of each component of the architecture. This architecture can be implemented in three practical scenarios: healthcare, call centers, and smart assistants.