End-to-end latency
Also known as：E2E latency · total latency · translation delay
End-to-end latency is the total time between the host speaking and an overseas viewer understanding the translation. Two definitions are common — speech to caption, and speech to the end of spoken translation — and they can differ by several seconds.
Two definitions
- Speech to caption — from the host’s first syllable to the translated caption appearing.
- Speech to end of spoken translation — from the host’s first syllable to the synthesised voice finishing the sentence.
The second is the moment the viewer has actually understood, and it is the one operational decisions should use.
What it is made of
End-to-end latency is the sum of five stages: segmentation wait, speech recognition, machine translation, speech synthesis and playback. Machine translation is usually the fastest of them; segmentation wait and playback dominate.
What it excludes
Delivery buffering is not part of end-to-end latency. OBS and the streaming platform each buffer, so the audience is already several seconds behind the host. That delay applies equally to original and translated audio, so it does not change how far the translation trails the voice — only the absolute wall-clock offset.
Include buffering when designing interaction such as answering comments; ignore it when chasing audio sync.
How to measure it
Start a local recording, say something with a sharp plosive (a three-two-one countdown works well), then measure in an audio editor from the start of the original to the end of the translation. Repeat ten times and take the median rather than the mean, because occasional long tails make averages meaningless.