What Sora 2 documents about speech

Sora 2 (developers.openai.com) is described as generating videos with synced audio, and its prompting guide is the one place in this register that tells a writer where a line of dialogue goes and how to label two people talking. From 3 source pages, read 2026-09-22; no maker pays to be listed.

Inclusion rule. Fields are the same five for every model here; a field the vendor does not document is recorded as not documented rather than inferred from generated output. Order. Fixed field order, identical on every model page.

1Dialogue has a place in the prompt, and that is unusual

Most entries here describe audio as an output and stop there. This one tells a writer where the words go: spoken lines sit in a dialogue block underneath the prose description, so the model can tell a description of a room from a line said inside it. Exchanges are asked to stay within a handful of sentences.

For a shot list that is a format instruction rather than a quality claim. It means dialogue in a production's own documents can be written in the shape the model expects, and the conversion step that usually sits between a script and a prompt mostly disappears.

2Arithmetic a writer can plan against

The guide puts one or two short exchanges in a four-second shot and a few more in an eight-second one, and warns that long, complex speeches are unlikely to sync. Read beside the 16 and 20 second generation lengths, and the extension that adds up to 20 seconds at a time to a total of 120, that is enough to size a scene before anybody spends a call on it.

It is also as close as the page comes to a statement about mouths: not a mechanism, but an admission of where alignment stops holding. Recording lip-sync as documented on that basis would be over-reading it, so the field stays marked unstated.

3Whose voice it is stays unaddressed

Nothing on these pages says where a voice comes from, whether one can be supplied, or whether a character keeps one between calls. Speaker labels route lines to the right face inside a single generation; a label is an instruction within one call, not an identity that survives the next.

For a serial that is the gap that matters. A production can write a two-hander that plays correctly in one shot and still have no published way to make the same character sound the same in the following episode.

4This entry, one column at a time

Each of these stays inside a single field: what Sora 2 puts there, what the wording settles, and what it leaves for a take to answer.

• Audio source — with the picture.
• Per-character binding — on the prompt.
• Languages — not documented by the vendor.
• Lip-sync — named only where it fails.

5Read against another entry

Each of these puts Sora 2 beside one other entry on a column where the two land at opposite grades of answer.

• Sora 2 and Luma Ray — on audio source.
• Sora 2 and Wan 3.0 — on audio source.
• MiniMax and Sora 2 — on lip-sync.

6Where this model sits against the rest

Inclusion rule. Models whose vendor documents something about generated speech. A model with nothing published on the point is still listed, with the cell marked hollow. Order. Alphabetical by model name.

Inclusion rule. Models whose vendor documents something about generated speech. A model with nothing published on the point is still listed, with the cell marked hollow. Order. Alphabetical by model name.

Inclusion rule. Models whose vendor documents something about generated speech. A model with nothing published on the point is still listed, with the cell marked hollow. Order. Alphabetical by model name.

7Sources

Taken from the model page and prompting guide at developers.openai.com on 2026-09-22. Every entry against the same five fields sits on the speech controls page; the rule behind each field is on how read. The entries either side of this one: sync-3, Synthesia.

• Audio source
  Described as a media generation model generating videos with synced audio, with video and audio both listed as outputgenerated with the pictureOpenAI, Sora 2 model page / recorded 2026-09-22
• Shot length a line has to fit
  Generations run 16 or 20 seconds, and an extension adds up to 20 seconds at a time to a total of 120OpenAI, video generation guide / recorded 2026-09-22
• How a spoken line is written into a prompt
  Dialogue belongs in a dialogue block below the prose description, with exchanges limited to a handful of sentencesOpenAI, Sora 2 prompting guide / recorded 2026-09-22
• Two characters in frame
  For multi-character scenes the guide asks for speakers labelled consistently and turns alternated, so each line lands on the right characteraddressed rather than skippedOpenAI, Sora 2 prompting guide / recorded 2026-09-22
• Dialogue timing against clip length
  A four-second shot is said to accommodate one or two short exchanges and an eight-second clip a few more, with long speeches unlikely to syncOpenAI, Sora 2 prompting guide / recorded 2026-09-22
• Languages
  Not documented by the vendor (as of 2026-09-22)neither a list nor a countOpenAI, video generation guide / recorded 2026-09-22