Kuaishou enhances live captioning speed and accuracy for e-commerce streams

Kuaishou has developed a new live captioning system that significantly reduces the delay between a host's speech and the display of captions. The end-to-end delay has been cut from 1.5 to 2 seconds to 400 to 500 milliseconds, with the character error rate dropping from 7.81% to 3.51%. This advancement supports millions of concurrent users, marking the first systematic engineering approach in e-commerce live streaming, which demands high concurrency and real-time performance. The system splits caption delivery into four stages: automatic speech recognition, PTS alignment, cascading distribution, and client-side rendering. The most complex part is the streaming revision mechanism, which continuously updates captions as speech is processed. This requires tracking sentence IDs, revisions, and player timelines to ensure captions are stable and do not flicker or overlap. This engineering approach is considered more valuable than focusing solely on recognition accuracy for teams building high-concurrency real-time AI systems.