AN APPLICATION FOR REAL-TIME DIALOGUE TRANSLATION AND SUMMARIZATION

Abstract

Real-time bilingual communication is frequently hindered by the high computational latency of modern Transformer models on consumer hardware. This thesis addresses this challenge by designing a real-time dialogue support system capable of Automatic Speech Recognition (ASR), Machine Translation (MT), text-to-speech conversion (TTS), and conversation summarization for Vietnamese and English. The inference engine was optimized using CTranslate2 with 8-bit quantization to enable efficient CPU deployment. Additionally, BART-base and BARTPho models were fine-tuned with adapters to handle bilingual summarization and headline generation. The optimized system achieved a Real-Time Factor (RTF) of approximately 0.6, successfully reducing the latency of the heavy Vietnamese pipeline to match lightweight English baselines. Evaluation results demonstrate that the proposed cascaded architecture outperforms an experimental End-to-End baseline, achieving a BLEU score of 38.98 and a latency of 3.95s, validating its suitability for real-time scenarios. Furthermore, ROUGE-1 scores of 49.91 (BART-base) and 58.15 (BARTPho) confirm robust capabilities in identifying main topics and participants. The study concludes that this optimized cascaded architecture offers a superior balance of speed, stability, and modularity compared to End-to-End alternatives, providing a viable solution for low-resource edge devices.