Transformer-Based Multilingual Simultaneous Interpretation System for Business Conferences

With the deepening of globalization, the demand for multilingual communication in business conferences is increasing, while traditional human simultaneous interpretation faces challenges such as high costs and scarce resources. This paper proposes a Transformer-based multilingual simultaneous interpretation system for business conferences (TMSIS), which employs adaptive wait-k strategies and domain adaptation mechanisms to ensure translation quality while effectively controlling latency. The system core comprises three modules: streaming speech recognition module, adaptive simultaneous machine translation module, and multilingual text-to-speech synthesis module. Extensive experiments were conducted on a multilingual parallel corpus in the business conference domain. Experimental results demonstrate that the proposed system achieves excellent performance across four language pairs: English-Chinese, English-Japanese, English-German, and English-French. Specifically, the system achieves a BLEU score of 35.920 on the English-Chinese translation task, with average latency controlled within 245.800 milliseconds, representing a 26.2% improvement in translation quality and a 13.8% reduction in latency compared to the baseline Wait-k model. Ablation experiments further validate the effectiveness of the adaptive strategy and domain adaptation mechanism.