Emotion detection represents a vital frontier at the inter-section of affective computing, artificial intelligence, and human-computer interaction (HCI). While traditional uni-modal systems—relying solely on text, speech acoustics, or static facial photographs—have established founda-tional benchmarks, they remain fundamentally fragile in real-world conversational dialogues characterized by lex-ical ambiguity, acoustic sarcasm, and visual occlusions. In this paper, we propose an end-to-end, publication-grade deep learning architecture for Multimodal Emo-tion Detection in Distributed Big Data Streams. Our framework synchronizes three specialized foundation models: a linguistic contextual transformer (RoBERTa-Large), a self-supervised acoustic representation encoder (Wav2Vec 2.0), and a spatiotemporal Vision Transformer (ViT-B/16) integrated with Facial Action Unit (FAU) tracking. To resolve temporal granularity discrepancies and inter-modal information imbalance, we formulate a novel Hierarchical Cross-Modal Attention Fusion (HCAF) network. The HCAF module leverages bidirec-tional multi-head cross-attention to project acoustic and vi-sual features into the linguistic latent subspace, selectively attending to emotional cues across communicative chan-nels. To address severe affective class skew (where neutral utterances dominate dialogue corpora), we design a hy-brid multi-task loss combining Focal Loss (γ = 2.0) with Concordance Correlation Coefficient (CCC) optimization for continuous Valence-Arousal dimensional estimation. Furthermore, we architect a distributed stream ingestion pipeline capable of sustaining over 450 concurrent mul-timodal streams per GPU node with sub-50ms latency, incorporating biometric facial landmark de-identification and vocal tract perturbation to guarantee user privacy under GDPR Article 9. We perform extensive empirical evaluations on the Multimodal EmotionLines Dataset (MELD) and the Interactive Emotional Dyadic Mo-tion Capture (IEMOCAP) corpus. Our HCAF model achieves a weighted F1-score of 71.6% on MELD (7-class) and 82.3% on IEMOCAP (4-class), outperforming uni-modal text, audio, and visual baselines by 7.7%, 27.5%, and 30.8% respectively, while surpassing standard Early and Late fusion paradigms by 5.9% and 4.6%. Ablation experiments and Pareto latency analyses confirm that cross-modal self-attention and dynamic loss re-weighting are crucial for robust, real-time affective computing.

 

Keywords: Affective Computing, Multimodal Emotion Recognition, Vision Transformer, Wav2Vec 2.0, RoBERTa, Cross-Attention Fusion, Focal Loss, Distributed Stream Analytics, Biometric Privacy Preservation.