Multi-level Feature Extraction for Bangla Speech Emotion Recognition

Emotion recognition from speech plays a vital role in enhancing human-computer interaction, mental health assessment, and social robotics. Despite progress in emotion recognition for widely spoken languages, Bangla remains underexplored due to limited datasets and unique language characteristics. Existing CNN models often fail to capture the nuances of Bangla emotional speech, largely due to the scarcity of diverse data and variations in acoustic properties. This study introduces a CNN-based model tailored for Bangla speech emotion recognition where multiple features were considered for better analysis, using the SUBESCO dataset and two more datasets known as KBES and BASER were used to test the model’s generalizability. We addressed data limitations by applying three distinct data augmentation techniques, improving the model’s robustness. Feature extraction focused on different level of key audio characteristics such as Zero-Crossing Rate (ZCR), Mel-Frequency Cepstral Coefficients (MFCC), Mel Spectrogram, RMS value and Short-Time Fourier Transform (STFT) to capture critical emotional cues. Our model achieved an accuracy of 92.87% on the SUBESCO dataset and nearly 90% on both KBES and BASER dataset, setting a new benchmark in Bangla emotion recognition. The proposed model provides a strong foundation for future research and highlights the potential for further advancements in Bangla speech processing, aiding applications in natural language processing and human-centered technologies.