"Latency" is three different questions
Most confused caption-latency conversations are two of these quantities being talked about as if they were one. Separate them and the numbers stop contradicting each other.
The three quantities
Latency, processing time, delay
Video latency
How far the outgoing video runs behind the incoming video: the base latency of the protocol plus whatever buffer is configured. Sub-second latency is achievable with any output type.
Processing time
How long the pipeline takes to produce a caption, subtitle or dub from the spoken word: speech recognition, then translation, then speech synthesis.
Delay
What the viewer perceives: the caption against the audio they are hearing. Delay is processing time minus latency, floored at zero.
The trade-off
The buffer is the control
Raising latency reduces delay. That is the whole trade-off, and it is tuned per deployment rather than fixed.
Zero buffer
The video passes through with only the path's base latency, and viewers see captions a moment after the words are spoken: sub-2s caption delay in SRT, and typically about 5 seconds behind the program audio in HLS workflows.
A few seconds of buffer
For same-language captions, a buffer of 3-5 seconds absorbs the processing time entirely and captions land with no perceptible delay, in sync with the audio.
Translation and dubbing
Translated subtitles and dubbed audio are delivered in real time, and need a larger buffer than captions to land in sync, because translation takes longer than transcription. The buffer is tuned per deal rather than quoted as one number.
Base latency counts too
Time the path already spends, like a segment duration on CMAF ingest, is time in which processing happens. Buffer and base latency overlap rather than sum, so total latency is closer to whichever of the two is larger.
Per path
Base latency by delivery path
What each path adds to the video before any buffer, and what that means for captions on it. Every stream path accepts added buffer except HLS-at-origin in append mode, which adds nothing and changes nothing about your stream.
- SRT / MPEG-TS
- The video path itself adds a half-second latency in SRT workflows: about 0.3 seconds of muxing plus each SRT leg. Sub-second caption insertion into SRT streams, with sub-2s caption delay; typical captioning solutions display live captions within 4 to 8 seconds.
- CMAF ingest (DASH-IF)
- One segment of base latency, which also counts toward absorbing caption processing.
- HLS at the origin (append mode)
- A true zero: the existing stream is augmented, not delayed. Viewers typically see captions about 5 seconds behind the program audio in HLS workflows, since there is no buffer on this path to absorb processing.
- Widget
- About one second of base latency, with captions typically 1-3 seconds end-to-end from spoken word to caption display.
Frequently asked questions
Can captions appear exactly in sync with the audio?
Yes. Configure a buffer of a few seconds so the video's latency absorbs the caption processing time, and captions land with no perceptible delay. Many customers instead choose zero buffer and accept a short caption delay, typically under 2 seconds in SRT workflows. Which is right depends on whether your audience tolerates delayed video or delayed captions better.
How fast are SyncWords captions compared to typical solutions?
Sub-second caption insertion into SRT streams, and sub-2s caption delay in SRT, where typical captioning solutions display live captions within 4 to 8 seconds. On the events widget, captions appear typically 1-3 seconds end-to-end from spoken word to caption display.
What is the latency for translated subtitles and dubbing?
Translated subtitles and AI dubbing are delivered in real time. Translation takes longer than transcription, so landing them in sync needs a larger buffer than captions do, and the right setting depends on your languages, quality settings and tolerance for video latency. We tune it with you rather than quoting one number.
Does adding captions slow down our video?
Only by the path's base latency plus whatever buffer you choose. On the SRT path the video itself adds a half-second latency; on HLS-at-origin in append mode it adds nothing at all, because the original segments are never touched. There is no transcoding on any path.
Why does SyncWords insist on three separate words?
Because a single "latency" number is usually an answer to the wrong question. A vendor quoting caption speed may mean processing time; an engineer asking about latency usually means the video path; a viewer complaining means delay. The three are related, buffer converts between them, but they are different quantities with different numbers.
Tune it on your own stream
The right buffer is a decision about your audience, not a spec. Run a proof of concept with our engineers and see the trade-off live on your own content.