What is Dialogue timing? Definition, key points and related tools
Converting a line into shot seconds by speech rate: ≈ 5 Chinese characters or 2.5 English words per second, +0.75 s per sentence-final mark. Too long and the model repeats or stretches; too short and the tail lands on the next shot.
Why the seconds, not the words, are the constraint
With native audio the line and the picture come out of the same call, and the call has a length fixed before generation: Veo 3.1 offers 4, 6 or 8 seconds (8 is mandatory when reference images are attached), Seedance 2.0 takes 4 to 15, Wan 3.0 2 to 30. The model has to spread your words over the seconds you paid for. Dialogue timing is the arithmetic that makes the two match, so each line starts and ends while its speaker is on screen.
It is done per shot, not per segment. A 12-second segment with 5 seconds of dialogue works when the lines sit in the shots where each speaker is framed; it fails when a line starts in the speaker's close-up and finishes over the listener's reaction (see segment vs shot).
Where 2.5 words per second comes from
People talk faster than that. In the Switchboard corpus of American telephone conversations, Yuan, Liberman and Cieri (Interspeech 2006) measured 196 words per minute overall, about 3.3 words per second, and 236 with silences removed, with individual conversations ranging from 111 to 291. Planning at 150 words per minute (2.5 per second) deliberately budgets below conversation, for reasons the same paper helps explain:
- Short lines are slow lines. Speaking rate rose steeply across utterances of one to seven words before levelling off, so a three-word line takes longer per word than a fifteen-word one. That is why duration tables give very short lines a floor instead of dividing by 2.5.
- Line endings stretch. The final word of an utterance was the longest. The 0.75 seconds per sentence-final mark is the planning allowance for that lengthening and the pause after it.
- Drama needs air. A performed line is followed by a lip close, a glance or a reaction, and those happen inside the same seconds. Conversational speed leaves no room for them.
Timing a shot, step by step
- Count spoken words as they will be said. A figure such as 1,500 is spoken as two to four words (fifteen hundred, one thousand five hundred); a contraction such as didn't counts as one.
- Divide by 2.5 and add about 0.75 seconds for each sentence-final mark beyond the first; the duration table's base values already contain one.
- Add the after-line beat. The storyboard template table already rounds it in: 8 to 9 words become a 6-second shot.
- Snap to the model's grid. On Veo 3.1 a 5-second exchange becomes a 6-second clip, or an 8-second one if references are attached; write what fills the difference.
- Check the subtitle. Netflix's English style guide allows up to 20 characters per second for adult programmes and 42 characters per line.
Worked example: You knew about this all along, didn't you? is 8 spoken words and 42 characters. Speech needs 8 / 2.5 = 3.2 seconds; reading needs 42 / 20 = 2.1 seconds and one subtitle line. Speech is the binding constraint, and the template's 6 seconds leaves about 2.8 seconds for the pause, the lip close and a reaction.
Common mistakes and their causes
- Using the wrong unit. Mandarin is planned in characters (about 5 per second), English in words. Applying the character rate to the English example above gives 42 / 5 = 8.4 seconds instead of about 3.
- Timing the segment instead of the shot. The total fits, but a line crosses a cut and lands on the listener's face.
- Forgetting forced durations. Attach references on Veo 3.1 and the clip is 8 seconds however short the exchange; unplanned seconds are the ones the model fills by repeating or stretching.
- Reusing English shot lengths for a dub. Translated text grows: IBM's guidelines, as cited by W3C (second-hand), put average expansion at about 130% for English strings over 70 characters going into European languages, and more for short strings. Re-time or regenerate localised segments (see subtitles and localization). Google's Veo documentation adds that languages other than English have not been evaluated for that model.
- Two speakers in one short shot. Give each change of speaker its own shot, so every line has a framed speaker and its own seconds.
Key figures and sources
| Figure | Value | Source | Date |
|---|---|---|---|
| Switchboard telephone conversations (Yuan, Liberman, Cieri 2006) | 196 words/min overall; 236 with silences removed; 111–291 across conversations | www.isca-archive.org | 2026-09-12 |
| Netflix English subtitle limits | up to 20 characters/s (adult), 17 (children); 42 characters per line | partnerhelp.netflixstudios.com | 2026-09-12 |
| Veo 3.1 clip lengths | 4, 6 or 8 s; 8 s required with reference images, 1080p or 4K | ai.google.dev | 2026-09-12 |
| Seedance 2.0 clip length (BytePlus ModelArk) | 4–15 s, default 5; dialogue in double quotes recommended | docs.byteplus.com | 2026-09-12 |
| Text growth in translation (IBM figure via W3C) | about 130% for English strings over 70 characters into European languages (second-hand) | www.w3.org | 2026-09-12 |
FAQ
Why plan at 2.5 words per second if people speak faster?
Measured telephone conversation averages about 3.3 words per second (196 per minute in the Switchboard study), but that figure contains no performance beats. A drama shot has to fit the line plus its lip close and reaction in the same seconds, and 2.5 words per second leaves that room. If one voice consistently sounds slow, raise the rate for that character rather than shortening every shot.
Does the same rate work for other languages?
No single figure does. Mandarin is planned at about 5 characters per second; for other languages, synthesize or record a few typical lines and time them. Kling lists Chinese, English, Japanese, Korean and Spanish for Kling 3.0 speech, while Google's Veo documentation says non-English languages have not been evaluated, so calibrate per model and language.
What should fill the leftover seconds?
A closing beat: a small action without moving across the frame, such as setting down a cup or looking away, or a silent reaction shot. Unfilled seconds are what the model fills itself, by repeating or stretching the line. The storyboard template shows the pattern.
Sources
- https://www.isca-archive.org/interspeech_2006/yuan06_interspeech.pdf
- https://partnerhelp.netflixstudios.com/hc/en-us/articles/217350977-English-USA-Timed-Text-Style-Guide
- https://ai.google.dev/gemini-api/docs/veo
- https://docs.byteplus.com/en/docs/ModelArk/1520757
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-api-reference
- https://www.w3.org/International/articles/article-text-size.en
- https://kling.ai/quickstart/klingai-video-3-model-user-guide