Hearing a voice before a season depends on it
Casting is a listening decision and almost nothing here documents anything to listen to. The substitute is a deliberate test: one short generation per character, heard and kept, before any episode is written around it. As of 2026-09-12.
| What to produce | Why | What it prevents |
|---|---|---|
| A held note or a long vowel | Exposes artefacts a short line hides | A voice that fails only on emotion |
| A line with a proper noun | Names are the reliable failure case | Renaming a character in episode nine |
| A short exchange in the target language | Register and idiom are undocumented | Discovering a market is unusable |
| The same line twice | Shows how much a rebuild drifts | A cast that changes voice mid-season |
Inclusion rule. Elements of a pre-production voice test, chosen to exercise the behaviours no vendor documents. Order. Alphabetical by element.
1A paid test is cheaper than a season
Four short generations per speaking character is a trivial cost against a series, and it answers questions that no page in this register answers: how the voice handles names, how it behaves in the target language, and how much it moves between calls.
Keeping the results beside the shot list matters as much as running them. A finding nobody can retrieve is a finding that gets re-discovered by whoever is on shift when a shot comes back wrong.
2The same line twice is the most informative test
Generating one line, then generating it again from the same inputs, shows how much a voice drifts on a rebuild. Where a voice is stored and reused, the answer should be nothing; where it is re-established from a clip, it rarely is.
That single comparison decides how much archival discipline a production needs. It is also the test most obviously missing from every vendor page, because it measures consistency rather than quality.
3Proper nouns break reads more often than sentences do
Initialisms, invented words, brand names and names from another language are the reliable failures. There is no way to know how one will be read except by generating it, and a character name is the worst possible thing to discover late.
Testing the proper nouns of a production before writing a season around them is a twenty-minute job. Renaming a lead in episode nine is not.
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.
- Auditioning a voice — who documents a preview
- Casting a voice — picking and holding one
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: Singing and read copy, First tests.