Simultaneous Translation

Simultaneous translation (SimulT) aims to generate translations in real-time, before the entire source input is available, balancing translation quality with minimal latency. Current research focuses on improving SimulT performance using large language models (LLMs) and novel architectures like decoder-only transformers and non-autoregressive models, often incorporating adaptive policies to dynamically adjust the translation speed based on the available source information. These advancements are driven by the need for efficient and accurate real-time translation in various applications, such as live subtitling, interpreting, and automated video dubbing, and are leading to improved evaluation metrics that better reflect user experience.