Start with the manuscript, not the voice
Most AI audiobook tools start at the wrong end: they ask you to paste text into a text-to-speech box and pick a voice. That works for a blog post. It falls apart across a 90,000-word novel with a dozen recurring characters, shifting points of view, and dialogue that has to sound like the same person in chapter 2 and chapter 28.
A production-grade workflow starts by understanding the book. Before any audio is generated, the source text is parsed and normalized so that characters, scenes, and narration are identified from what the manuscript actually says — not guessed at line by line. In Marvox, this is the upload-and-analyze step, and everything downstream depends on it.
Build character intelligence from the source
Once the manuscript is understood, the next step is character intelligence: who speaks, how often, in what scenes, and with what relationships. This is the layer that keeps a character's voice consistent later, because a voice is only as stable as the identity it is bound to.
Marvox calls this layer CharacterOS. It builds a canon-grounded profile for each character from the text, which you can review before you spend time or money generating audio. Reviewing characters first is the single biggest lever for a good result — fixing a misattributed line here costs seconds; fixing it after audio is rendered costs a re-render.
Review scenes before you produce audio
Scene review is where you confirm cast, continuity, and direction before committing to production. Treating scene generation as a reviewable surface — rather than a one-shot export — is what separates a usable audiobook from a demo.
At this stage you are checking the cheap things while they are still cheap: is the right character assigned to each line, does the emotional read match the scene, are point-of-view shifts handled. Only after that do you move approved work into audio.
Produce multi-voice audio and export
With characters reviewed and scenes approved, audio production assigns voices per character and renders the result. Because the voices are bound to reviewed character identities, the same character reads consistently across the whole book.
The final step is export — moving the finished artifact out of the studio in a usable format. The throughline of the whole workflow is that understanding comes before generation, and review comes before production.