Simultaneous Translation at Baidu

Human Interpreters

Human interpreters face immense challenges in simultaneous translation due to the mental demands of processing and translating languages in real-time. Liang Huang explains that only 2,000 to 3,000 qualified interpreters worldwide can perform this task, and even they can only sustain it for about ten minutes before needing a break 1. This is because interpreters must strategically choose to cover only 60% of the source material to keep pace with the speaker, as attempting to translate everything would lead to significant delays.

Simultaneous interpretation is extremely difficult for human beings mentally. It's extremely challenging to concurrently comprehend one language in a headset and also speak in another language.

--- Liang Huang

This limitation highlights the need for technological advancements to support human interpreters in maintaining accuracy and efficiency 1.

   

Technology Limitations

Machine translation technology faces unique challenges, particularly with language order differences. Liang Huang notes that languages like German and Chinese have complex word orders that complicate simultaneous translation for machines 2. This complexity often requires machines to predict or wait for certain words, impacting translation accuracy and latency.

There's a trade-off between latency and quality. If you wait longer, you can predict better and translate better.

--- Liang Huang

Baidu's recent advancements in simultaneous translation aim to address these issues, achieving a delay of only about three seconds, similar to human interpreters 3.