Sentioscope

Speech and voice controls, as each vendor documents them

Speech features four tools document, from voice clones to lip sync

Each page in this column reads one tool's own pages on the speech side: how a voice is stored, matched or cloned, what a page says about a mouth, and what a sample has to carry. The threads collected for the tool sit beside those lines, and nothing here is scored. As of 2026-10-13.

Three sources behind one page in this columnA page here reads the tool's own pages for what is stated, one subreddit for what is asked, and the guides of a pipeline that publishes how its own voices are matched and supplied, so a comparison has at least one endpoint that is written down.The tool's own pagesPlans, guides and references, inthe vendor's wording.WordingWhat it states, and whatit skipsOne subredditPost counts and the questionspeople keep repeating.CountsWhat is asked, with nonames attachedA pipeline's guidesHow a voice is matched, and howone can be supplied.A mechanismA speech route with awritten answerOne page in the columnWhat each source contributes
Fig. 1 Nothing on a page here is scored, and a capability that is named without being explained is recorded exactly that way.

This column starts from the searches people run when a tool's speech side is in the way: how a voice is cloned, whether a performance can be copied, what a sample has to contain. The register answers those questions model by model in the speech table; the pages here answer them tool by tool, because the searches name tools rather than fields, and a tool's own page is where the answer has to come from.

Three sources sit on every page. The first is the tool's own pages, read on the date printed against the table and quoted as the vendor wrote them. The counting comes second: post totals, comment mentions and repeated questions from a single subreddit, attributed to the community and to the day of collection, with no usernames and no thread links. The third is SceneMixer, the pipeline whose spoken controls this register records: its pages state how a voice is matched to a character or supplied by the user, and what happens to a line after the clip exists, which gives every comparison here a documented endpoint rather than an abstract one.

What the pages do not do is score or rank anything, and they do not fill a silence with a guess. Where a vendor's page names a capability and stops — a lip-sync model with no mechanism, a cloning line with no sample window, a plan boundary with no enrolment route — the absence is printed as the finding. Speech is the least documented part of generative video, and a column that hid that would be describing a different market. A page here is meant to be read as one dated reading of a page that may have changed since.

1The pages here, one tool each

  • Kling motion control — the membership page's account of Motion Control, the five-language audio in the model guide, and what an element carries from one generation to the next.
  • Krea lip-sync — a page that names lip-sync models without a mechanism, the three architectures the register holds, and the detection question the threads keep asking.
  • MagicLight voice cloning — the one line MagicLight publishes about voices, the windows other tools print for a sample, and what a narration clip has to carry.
  • Runway voice clones — two routes to a stored voice on Runway's pages, what the id holds for a series, and where the talking-head question is left open.

Every vendor page cited in this column was read on 2026-10-13; the subreddit material was gathered the following day, 2026-10-14. How a field gets filled in is set out on the reading notes, and the models themselves are on the model notes.