Memory-Augmented Tone-Adaptive Multilingual Dialogue Generator: Real-Time Cross-Lingual Style Preservation via Neural Prototype Fusion

Multilingual dialogue systems often struggle to preserve the emotional tone and stylistic intent of messages during real-time translation and response generation. This paper presents a memory-augmented tone-adaptive multilingual dialogue generator designed to address this limitation without requiring per-language fine-tuning. The proposed approach combines a transformer-based architecture with a neural memory bank that stores cross-lingual style prototypes. During inference, the system encodes input utterances into tone representations, retrieves relevant stylistic patterns using a sparse attention mechanism, and integrates them into the generation process through a gated fusion strategy. To further improve consistency, an online calibration module adjusts the generated output in real time, reducing mismatches between intended and produced tone. By relying on shared stylistic representations, the system generalizes effectively across languages, including low-resource scenarios. Experimental results indicate improvements in both tone preservation and computational efficiency when compared with existing approaches. The framework offers a practical solution for multilingual applications that require consistent and context-aware communication across diverse linguistic settings.