What to find out before a platform gets chosen
Every column in this register has a version that only a take can answer. A short list of deliberate tests, run before a platform is chosen, settles more than any amount of reading, and one entry publishes a free allowance to run them in. As of 2026-09-12.
| Question | Cheapest test | What a page will never tell you |
|---|---|---|
| Does a rebuild drift | The same line twice, same inputs | How stable a voice is |
| Does the market work | One exchange in the target language | Whether idiom and register land |
| Does the two-shot hold | One wide with two faces, one talking | Which mouth a model will choose |
| Is the audio keepable | One shot, listened to on speakers | Whether a mix has to be rebuilt |
Inclusion rule. Questions raised by the columns in this register that can only be answered by generating something. Order. Alphabetical by question.
1Four tests, and they are the same four everywhere
The interesting questions do not change between products, which means a production can build one small test battery and run it against everything it is considering. That is more useful than a feature comparison and much faster than it looks.
The results belong in the notes a production keeps, with a date, because they describe a build on a day rather than a product. That is exactly why they are not in this register.
2A published free allowance is worth more than it sounds
One entry documents one generation a month with a fifteen-second ceiling. That is small, concrete and enough to answer the question its column raises: whether alignment is acceptable on footage the production made itself.
Most platforms have some trial arrangement and do not document it, which means a team evaluating them cannot plan the evaluation. Publishing the allowance is a small act of documentation with a large effect on a procurement schedule.
3Test the proper nouns before anything else
Names are the reliable failure case for synthesised reads, and a character name is the worst possible thing to discover late. Generating the cast list as spoken lines takes minutes and occasionally changes a script for the better.
It is also the test that transfers least between products. A name a model handles well is a fact about that model, so the test has to be repeated wherever a series might move.
4Where the published detail is logged
What is described above is how the techniques work in general. Which models document which of them, in whose words, is kept in the speech controls table with a date on every field.
- How a field gets filled in — why none of this is in the table
- Hearing a voice first — the pre-season test in detail
A mechanism note rather than a documented field. Nothing here is attributed to a model, and nothing here is a claim about one. The sourced material is on the speech controls table. Related: Subtitles and length, Asking a vendor.