Emotion recognition is a foundational task in affective computing with applications in human-computer interaction, mental health monitoring, intelligent assistants, and personalized services. Traditional unimodal approaches commonly fail to account for the subtlety involved in human emotional expression, particularly in scenarios of natural conversation. To this end, we have proposed a Deep Learning-Based Emotion Recognition System integrating text and audio modalities through an optimized multimodal architecture. The proposed model incorporates BERT-based text embeddings for deep contextual modelling, while the audio features are extracted through CNNs from Mel spectrograms to capture prosodic and tonal variations. These features are fused through a multi-head attention mechanism that enhances cross-modal alignment and representation. More importantly, the model is designed to allow federated learning for privacy-preserving deployment in distributed settings such as IoT networks. In evaluation on the MELD dataset, the proposed system recorded a weighted average accuracy of 65.86% and an F1-score of 65.12%, outperforming several existing multimodal emotion recognition methods.