Sentioscope

Speech and voice controls, as each vendor documents them

The speech notes as one file

The notes are exported as one file. Each row records what a vendor documents about generated speech, with the page it was read from. As of 2026-09-12.

A file where the empty cells are most of the findingThe same speech fields are kept for every model. Most models document the first and leave later ones unaddressed, so the file carries many rows recording that nothing is published. That imbalance is the most informative thing it contains.Audio sourceWhether soundcomes out withthe picture.Usuallydocumented.Then thinnerReference voiceWhether asample cansteer thespeech.Documentedunevenly.Then thinnerLip-syncWhat the vendorsays drives themouth, if itsays anything.Then thinnestLanguagesWhich languagesare named fordialogue ratherthan for menus.
Fig. 1 Exporting only the filled cells would describe a market where speech is well documented, which is not the market these notes found.

Speech is the least documented part of generative video, so this file is unusually full of rows recording that nothing is published. That imbalance is the most informative thing in it.

Each row names the field it fills, so a reader rebuilding the comparison cannot count a reference-audio claim as a native-audio one. Those two get merged constantly in product copy and the merge changes what a production has to plan for.

Nothing here comes from listening. Every row is a documented statement, which means the file describes what vendors commit to rather than what their models sound like.

Reuse under the licence below is permitted, including commercially, provided the source and date columns stay attached to the rows.

1What each column holds

One line of the export, explained. Recorded 2026-09-12.
ColumnWhat it holds
fact_idThe row's handle. Rewriting a note does not change it.
valueWhat the model's documentation states about speech.
applies_toThe control or language list the row is about.
source_urlThe documentation page it came from.
source_nameThe vendor's own name for that page.
value_sinceFrom when the documentation has said this.
verifiedThe last day that page was read.

Inclusion rule. Every recorded field is exported, including the many that record no published control. Order. File order follows recording order, grouping rows by model.

The file is at /datasets/controls.csv, licensed CC BY 4.0. Attribution should name the notes and the reading date.

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

2Voices in a drama pipeline

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

It is applied when the script is read, so changing it means re-reading. Video already shot keeps the lines it was shot with.

SceneMixer, Working in Your Own Language — “Can I change the dialogue language after generating?” (read 2026-10-13)

It is a property of the character, written once in the voice description, and it covers every line they speak. What changes line by line — mood, pace, a word under stress — goes in the delivery note in brackets at the start of that line.

SceneMixer, Give a Character an Accent or Dialect in an AI Drama — “Do I have to repeat the accent on every line?” (read 2026-10-13)

The quote column. Each reading in the file also carries the words on its source page that it was read from, so any row can be checked against that page. Rows recorded as not stated on the pages read leave it empty.