Emotion Recognition from Bangla Dialect Speech using Privacy-Aware Deep Learning Models: A Comparative Analysis

Speech emotion recognition (SER) is critical for creating affective and context-aware human-computer interaction systems. However, SER research in low-resource languages like Bangla is still restricted, especially in terms of dialectal variety and privacy-preserving model training. This paper introduces a Bangla dialect-sensitive, privacy-aware SER framework capable of recognizing five distinct emotional states: neutral, happy, sad, angry, and surprise.We study three hybrid deep learning architectures: a composite EfficientNet-Vision Transformer (EfficientNet-ViT) model, CNN-BiLSTM for extracting spatial-temporal patterns where EmoDARTS using differentiable architecture search for automatic optimization. With a 93.0% F1-score and 95.9% accuracy, EfficientNet-ViT outperforms the other models in a federated learning setting while maintaining data security among dispersed devices. To address data scarcity and improve model generalisability, we use a cross-lingual transfer learning technique. Models are pretrained on high-resource English SER datasets (RAVDESS, SAVEE, and TESS) and then fine-tuned on Bangla datasets (SUBESCO, BanglaSER, and a freshly generated dialect-rich corpus). The suggested technique efficiently handles the issues of dialectal diversity, resource constraints, and privacy in Bangla SER. This methodology shows great promise for scalable implementation in real-world applications and offers a reproducible blueprint for SER in other low-resource language environments.