Transcription is **asynchronous and non-blocking**: the caption node is on the canvas from the moment the project mounts and its transcript attaches once ready. Because it reads the scene's audible mix, it waits until **every other source in the scene has landed** — a generated `voice` or `audio` track is transcribed at its final placement, not at the placeholder. The scene must contain an unmuted, unhidden audio or video source; without one the node carries an error saying so, recorded in the library like any failed generation (see [errors.md](./errors.md#failed-sources)).
Give `src` a transcript file — `.srt`, `.vtt`, or a transcript `.json` — and it is mounted as it is; no transcription runs and no credits are spent. It resolves like any other [`src`](./media.md) (library path, asset id, path, URL), except that `generate.*` is not accepted.
| `verticalAlign` | `"top" \| "center" \| "bottom"` | preset default | Vertical placement of the caption block: anchored to the top or bottom safe margin, or centered. Horizontal placement stays with the preset. |
| `seed` | `number` | `0` | Joins the transcript cache key — see [Caching](#caching). Ignored when `src` is set. |
| `offsetX`, `offsetY` | `number` | `0` | Render-time nudge in px on top of the preset placement; subpixel values are kept. A slide animation drives the same channel and wins while it plays. |
| `start`, `end`, `sourceIn`, `sourceOut` | `Time` | full transcript | Trim which stretch of the transcript is captioned, using the same [timing](./timing.md) semantics as media nodes. See [Trimming](#trimming). |
| `id`, `name` | `string` | see [elements.md](./elements.md#common-props) | Address and label. |
`<animation>` children are valid; the text-only types (`"appearWord"`, `"appearChar"`, `"scramble"`) apply.
The preset positions the caption block; `verticalAlign` overrides only its vertical anchor (`whisper` and `cascade` default to `bottom`, all other presets to `center`), and `offsetX`/`offsetY` nudge the drawn result from there.
A transcript is cached under **the scene's id and the `seed`**, and the cached asset is reused whenever that pair comes up again — so reopening a project transcribes nothing and consumes no credits.
The audio itself is not part of the key. **Recutting a scene does not re-transcribe it**: to pick up changed audio, bump `seed`. A value used before replays that take from cache; a value that has not been used for this scene transcribes it again, which costs credits.
A transcription that failed is recorded under the same key, as a library entry in error, and is not run again until that entry is removed — see [errors.md](./errors.md#failed-sources).
Captions carry the same [timing](./timing.md) props as media nodes, and the transcript is source content that must stay aligned to the audio — so advance `start` and `sourceIn` together (and `end`/`sourceOut`), never `start` alone. To show only from 15 s onward, set both `sourceIn={15}` and `start={15}`. One node can't skip a gap, so to blank captions out for a middle stretch — e.g. under an overlay — use two nodes that meet at the gap:
`preset` selects the caption style: the same presets as the editor's caption inspector. Some presets expose **color slots**, filled in order by the `colors` prop; a missing or omitted entry falls back to the slot's default.