Transcription and speech evidence
Transcription converts an audio track into words and timestamps. It supports tasks such as Video Analysis and subtitles. It does not generate a voice or establish that a speaker said a phrase correctly just because the recognizer returned plausible text.
Two Studio uses
Section titled “Two Studio uses”Detailed Video Analysis uses speech recognition alongside visual analysis. Quick/style analysis does not run this speech-recognition path and should not be used as dialogue or speaker evidence. Its primary speech-analysis path can provide speaker information. A fallback can lose speaker labels; treat a visible degraded result as incomplete evidence, not as proof that only one person spoke.
Subtitles use their own transcription path. Their input «звук/видео» takes either a video clip or an audio track (a voice-over, a dubbed track, an uploaded sound file). Do not assume that subtitle timestamps, speaker labels and Video Analysis have identical model selection or fallback behavior.
The Subtitles node has a recognition model field (since 18 October 2026) with two choices:
- Scribe v2 — songs and music (ElevenLabs), the default of every Subtitles node, including nodes saved before the field existed. It returns every word with its own timing, so Studio builds each subtitle line from the words: a new line starts at a pause, at the end of a sentence or when the speaker changes, and a line begins with its first word. Sound events that Scribe tags, such as
[music]or(laughter), do not become subtitles. - Whisper large-v3, an explicit choice. A node that already has a result keeps it and does not become stale or re-run by itself; its status line names the engine that made that result, and the next run uses the selected model.
Why Scribe is the default: on songs Whisper can collapse a 30-second window into a single word. In a production project on 18 October 2026 Whisper returned one word for the opening rap and jumped to 0:30, so seven lines were missing; Scribe v2 recognised the same rap from 11.22 s with word timings. This is one observed song, not a benchmark — still read the result. If Scribe v2 is unavailable, Studio recognises with Whisper large-v3 instead and the node’s status line names the engine that was used and why (for example transcribed by Whisper large-v3 (Scribe v2 unavailable: HTTP 401 …)); if both fail, the run fails with both reasons. The price of a Subtitles run is the same for both engines (see the estimate before the run).
Example and verification
Section titled “Example and verification”Analyze a short clip with one clearly audible speaker. Compare the transcript against what you hear, especially names, numbers and the final word. Inspect whether timestamps fit the spoken phrases. With multiple speakers, verify each attribution and check overlaps manually before using the analysis to write another scene.
Silence can produce a false transcript; a plausible transcription can miss a mispronunciation. In the 11 October 2026 HappyHorse speech trial, the transcript looked largely reasonable while the spoken word still needed correction. Treat text, listening and visual lip movement as separate observations.
With Whisper, subtitle timing uses the word-level layer of the same recognition: a subtitle line starts at its first spoken word (never earlier than the recognised segment), while its end and text stay as recognised. Re-run the Subtitles node with 🔄 to apply this to an older project (about $0.01 for recognition plus about $0.02 for a new Decorate plan; see the estimate before the run). If the recognizer returns no word timings, the lines keep the segment timing. The word timings also travel with the VTT to the Decoration node (as WebVTT word marks), which uses them for its TikTok look; downloaded SRT and VTT files do not contain the marks.
The current service code accounts for Scribe v2 and Whisper-family recognition paths. Their availability does not promise equal speaker labeling or punctuation. Engine identifiers and fallbacks belong to the service contract; the only recognition picker in the UI is the Subtitles node’s recognition model field described above.
Reviewed 18 October 2026 against the analysis and subtitle contracts, the dated speech trial and the 18 October 2026 song comparison. No new recognition benchmark was run. See dialogue, subtitles and decoration and MCP support.