Where the sound in a generated shot comes from

Sound divides into speech generated with the picture, audio laid over it afterwards, and nothing published at all. Two entries reach the top by documenting generated speech in named languages with voice selection; three sit at the bottom with output that is silent or undescribed. As of 2026-09-12.

Inclusion rule. Tools rated on this rubric: a public English product page addressing at least four of the seven criteria. A tool the vendor documents too thinly to rate is left out rather than rated low. Order. Descending on this criterion, then alphabetical; unrated last.

1Three different things called audio

Generated speech that arrives with the shot, a soundtrack added in an editor, and an uploaded track driving a mouth are three mechanisms with three different failure modes. Vendors describe them in the same paragraph often enough that the distinction has to be drawn by the reader.

The rating draws it: speech generated with the picture, in languages the vendor names, sits at the top, because that is the one a serial production can plan a cast around.

2Named languages are what lift a rating

An entry claiming multilingual speech without listing the languages cannot be planned against, and is rated as a partial statement. A named list is checkable, and it is the difference between a market decision that can be made now and one that waits for a test.

This is the same principle the rest of the rubric runs on, and sound is where it bites hardest, because language coverage is usually the first question a distributor asks.

3Voice identity across episodes gets almost no coverage

Nearly every published statement here is about generating a voice once. Keeping the same voice across forty episodes is addressed by very few, and where it is, the mechanism is usually a preset library rather than anything stored per character.

The ratings cannot capture that on their own, so it is worth reading the sound column beside the cast consistency one: a recognisable face with a drifting voice fails in the same way a drifting face does.

4Lip alignment is documented least of all

Of the three mechanisms, how the mouth is driven is the one vendors describe least. Speech generated with the picture usually implies alignment and rarely states it; audio laid over afterwards usually does not have it and rarely says so. A reader planning close-ups has to test rather than read.

That gap is why this column tops out at five rather than being split further. There is not enough published material across the field to rate alignment on its own without inventing a distinction.

5The floor here is mostly silence

Entries at two publish output with no audio statement at all. That is not a defect — several are explicitly picture-only tools — but it moves the whole sound budget outside the rate card, and it should be counted before the rate is compared.

6Cells in this column with a page of their own

Values in this column held by at most two of the rated tools, each read against the rule above: Hailuo, sound, Veo and Flow, sound. The crowded middle of a column gets no entry, because a value seven tools share describes the field rather than any one product.

Ratings describe what each vendor publishes, read 2026-09-12, and are given under the rule on how scoring works. All seven criteria together are on the ratings table. Next: Assembly, Cost transparency.