Speech questions with answers in the documentation

A question is answered here when documentation settles it. How a voice sounds, whether a performance is convincing, and whether a listener would notice a seam are all outside what a page can establish. As of 2026-09-12.

Fig. 1 The right-hand column is where most of the value of a speech feature sits, and no page can establish it.

The exclusion is unusually painful for this subject, because speech is judged by ear and the ear is exactly what a register cannot use. A note can say that a model accepts a reference voice; only listening reveals how closely it follows one.

What documentation does settle turns out to shape a production anyway. Whether audio arrives with the picture decides whether a dubbing stage exists. Whether a language is named decides whether a market is reachable. Whether timing is controllable decides how the edit is planned.

Answers link to the notes, and the notes link to the vendor pages, so a claim can be traced to the sentence a vendor actually published.

1The questions

Keeping a voice — Five models keep a voice as a stored object you approve once.
Lip-sync — One vendor ties it to the speaker on screen, one drives it from the script, one names the feature without describing it, and one removes the step.
Voice sources — A catalogue, a clone from a recording, a finished track, or a voice described in words.
Sound by default — Two vendors state that sound comes back unless it is switched off, and one of them publishes what switching it off does to the price.
Which language — One vendor publishes a catalogue with language codes, one names fifteen dialogue languages, five publish a count or a bare field, and nine publish nothing.
Two in frame — Two characters in frame with a line each is the commonest shot in drama, and exactly one vendor publishes any guidance for it.
Who docs are for — Developer references describe parameters.
How long a shot can be — Six entries publish a length for a generation or its audio, from a four-second shot to a ten-minute talk.
Fifteen seconds of voice — Two entries cap reference audio at fifteen seconds in total, and the published windows elsewhere run from one recording to twenty minutes of studio time.
Answers software can check — Most of this register is prose.
What silence costs — One entry states plainly that turning audio off does not change the rate.
Where a line goes — Four entries tell a writer where dialogue belongs: a labelled block, a clause after one keyword, a quoted line, or a reference voice instead of words.
Auditioning a voice — Two entries document hearing a voice before generating with it: a preview on a ready voice, and a sample for every preset in a library.
Outside suppliers named — Two entries name outside companies for speech: one lists five providers behind its voice field, and one names an audio house as a planned integration.
The column always filled — Audio source is filled by fourteen of the sixteen entries.
Speech without the bed — Two entries document dialogue without the music: a speech-only mode, and a switch on a score added later.
Regenerating a shot — Where audio comes out with the picture, redoing a shot produces a new performance.
Consent in the docs — One entry treats cloning an identifiable voice as a likeness question and names digital replica statutes.
A second control — Four entries publish a control most of the register lacks: a pose track, a speech-only mode, a reverse audio route, or a voice asked for in words.
What kind of page — The shape of an answer follows the kind of page it sits on: an API reference, a model card, a prompting guide or a compliance checklist.
Formats and file sizes — Four entries publish what they accept as a file: WAV or MP3, a longer list of containers, or a ten-megabyte ceiling on a voice sample.
What sets the length — On the pages read, length comes from two places: a menu of fixed options or a range a caller chooses, with extensions stitched on in one case.
Charging for audio — No entry here bills audio as a separate line.
Asking for an accent — Three entries document asking for an accent: in a voice-design prompt, in native audio, or in a character's voice line that an attached voice overrides.
Changing a line — Four entries document re-voicing finished video: a lip-sync pass that keeps the picture, a re-performance, or a dubbing project on an uploaded cut.
Second language — Three entries dub finished video into other languages.

Where documentation is silent, the answer says which sentence would have to appear for the question to become answerable.

Other notes: Speech controls, Models, Fields, Routes, Side by side, Terms, Learn, Data. What counts as a documented control is set out on the reading page.

2How one tool handles speech

For the speech side of “Questions”, SceneMixer's guides describe how lines are spoken:

The spoken lines and the end card follow your interface language across fifteen dialogue languages, with the cast and setting adapted to that market. The end card falls back to a language it can typeset when a script needs shaping or right-to-left layout.
SceneMixer, AI Ad Video Generator — “Which languages can the ad be in?” (read 2026-10-13)

SceneMixer writes dialogue into the script analysis, assigns each line to a shot with a speaker and a short delivery note, and the video model performs it — picture and audio generated together, so the mouth forms that sentence in the language you chose.
SceneMixer, AI Dialogue Generator (read 2026-10-13)