Changelog
Release notes
Everything we ship, in order. The writing engines, the Co-writer, the studio, and the production tools that carry a script or a book through to screen.
One audio stack
The ad hoc speech code was replaced by a single audio module built on the current ElevenLabs SDK, covering synthesis, script speech, transcription, music, sound effects, voice changer, audio isolation, voice design, voice cloning, and dubbing behind one client, one policy, and one cache.
The same set of audio controls — voice picker, voice chip, settings, audio tags, clone and design dialogs — is now reused everywhere: text-to-speech, casting, talent, the dialogue tabs, the timeline's generate-audio panel, and book chapter listening.
Voices, at the project level
The per-episode voice studio is gone. Voices now live at the project level, where casting actually happens — one page listing your voices with previews and per-voice actions, serving every episode under the project.
- Suggested voices: a character carries a suggested voice until you assign one, so a cast has a sound before you have cast it
- Delivery cues: direct a line's performance server-side rather than hoping the punctuation carries it
- Custom voices are owned. A cloned or designed voice belongs to the account that made it, and only that account can use it.
Books you can listen to
Book chapters play with read-along, the text following the audio as it goes.
Dubbing that settles honestly
Dubbing is tracked as a real job with server-measured source duration, so it bills once, for its actual length. If the process dies between submitting a dub and recording it, an hourly sweep now finds the orphan and cancels it — instead of a job that runs, costs, and belongs to nobody. Finished dubs are never touched.
Priced from the real rate card
The single flat audio charge became nine per-capability features, each priced from the live model and voice catalog rather than a fixed cents-per-minute row. Speech and voice-training runs place a cost hold before they dispatch, settled by the meter that follows.
Media generation, tightened
- Provider results are read through the provider's own contract. A result that does not match is held or refused rather than billed on the strength of a claim.
- The catalog was refreshed and every newly carried model is priced before it is offered
- References are taken as ids the server resolves, not URLs the client sends
- Media is fetched only from hosts we have named
- Late-settled rows are re-checked for cost, and ledger discrepancies are reported rather than absorbed