Sora 2: a label routes a line, and then it is gone
Speaker labels inside the prompt send each line to the right face. A label is an instruction within one call rather than an identity that survives the next, so nothing here makes a character sound the same in the following episode. As of 2026-09-22.
| What the documentation settles | What it leaves to a take |
|---|---|
| Speakers should be labelled consistently, with turns alternated | Whether the same label produces the same voice twice |
| Dialogue belongs in a dialogue block below the prose description | Whether a label can be tied to anything outside that block |
| Nothing is published about where a voice comes from | Whether a voice can be chosen, supplied or reused at all |
Inclusion rule. Only statements found on the vendor page named in the caption, and only ones bearing on this one field. Nothing here is filled in from generated output. Order. In the order the statements arrive on the vendor's page.
1Labels solve the shot and not the series
Inside one generation a label does real work: it tells the model which of two people is speaking, which is the distinction most vendors leave to chance. The guidance is careful about it, asking for consistency and alternating turns.
Across generations it does nothing, because there is no stated mechanism connecting a label to a voice. A two-hander can play correctly in one shot and come back with different voices in the reverse angle.
2Consistent labelling is still worth doing
Even without persistence, labelling the same character the same way keeps a shot list readable and keeps the model's behaviour inside one call predictable. It also means that if a voice mechanism ever arrives, the production's documents are already in the shape it would want.
What it cannot do is substitute for a cast. A serial needs voices that outlast a call, and this entry publishes nothing that does, which is the finding rather than the workaround.
3The others that re-establish the voice on each call
Four more entries put the voice on the request. All four pass something that identifies a voice; this one passes a name that identifies a speaker and leaves the voice unaddressed.
- Two characters in frameFor multi-character scenes the guide asks for speakers labelled consistently and turns alternated, so each line lands on the right characteraddressed rather than skipped
- How a spoken line is written into a promptDialogue belongs in a dialogue block below the prose description, with exchanges limited to a handful of sentences
- LanguagesNot documented by the vendor (as of 2026-09-22)neither a list nor a count
4Sources
Read from the model page and prompting guide at developers.openai.com on 2026-09-22. The same column across every entry is on per-character binding; everything this vendor publishes about speech is on Sora 2. What counts as documented is on how read.