Abstract

Speech Emotion Recognition (SER) is a challenging task because tone, pitch, and speaking habits vary widely among individuals. Machine learning models often struggle to generalize in such heterogeneous scenarios. This paper presents a robust SER model that combines digital signal processing (DSP) techniques — namely MFCC, spectrogram, and chroma feature extraction — with deep learning architectures such as CNNs, RNNs, and Transformer-based models. An advanced hybrid CNN-LSTM model is implemented to effectively capture both spatial and temporal features, leveraging the complementary strengths of each architecture. Robust preprocessing and data augmentation techniques are used to enhance the model’s resilience to speaker variability and noise. Experiments conducted on the RAVDESS dataset evaluate the comparative performance of various models using key performance metrics, including accuracy, precision, recall, and F1-score. The experimental results demonstrate that the hybrid CNN-LSTM approach significantly outperforms traditional and standalone deep learning methods, achieving higher accuracy and better generalizability across a wide range of emotional expressions.