The Kuaishou E-commerce Live Streaming Technology Team has introduced a real-time captioning engineering paradigm called "Live Captioning Engineering," advancing the live captioning process from simply "having text" to "real-time and accurate text." The most direct figures are: the end-to-end delay from the host's speech to the caption being displayed has been reduced from 1.5 to 2 seconds to 400 to 500 milliseconds, and the character error rate (CER) has dropped from 7.81% to 3.51%. Behind this is also the ability to support millions of concurrent users—this is the first systematic engineering approach in the e-commerce live streaming scenario, which is known for high concurrency and strong real-time requirements.

The core of this paradigm is to split the delivery of captions into four consecutive stages. The first stage is automatic speech recognition (ASR), where the model converts the host's audio stream into text in real time; the second stage is PTS alignment, which aligns the recognized text with the timestamps of the audio and video to ensure that the captions match the speaker's mouth movements; the third stage is cascading distribution, where the recognition results are sent through a service pipeline to millions of viewers; and the fourth stage is client-side rendering, where the player on the user's phone displays the captions. These four stages connected together complete the "speaking—displaying text" loop, and any delay in any stage will directly affect the latency seen by the audience.

The most challenging part is the streaming revision mechanism. Live speech is continuous, and the model revises as it listens, so the recognition result of the first sentence may be completely revised when the second half is heard. Therefore, the system must be able to distinguish "which sentence, which revision, and where it should be placed on the timeline"—this is exactly the purpose of the trio of sentenceId, revision, and player timeline. sentenceId locks a complete sentence, revision marks the version of the sentence, and the player timeline determines at which frame the revised caption should be updated. Only by working together can they stabilize the continuously revised captions on the screen without flickering or overlapping. For teams looking to build a high-concurrency real-time AI system, this engineering experience that thoroughly explains "streaming, revision, and alignment" is more valuable than just focusing on recognition accuracy.