Reply to thread

Message: [QUOTE="Bombastus, post: 91172, member: 2178"] ElevenLabs says its audiobook character detector can miss minor, unnamed, or otherwise ambiguous speakers in an uploaded manuscript. Automatic casting is therefore a useful first pass, not a reliable record of who speaks every line. The usual workflow makes the risky part look easy. A tool finds names, proposes voices, and previews dialogue while the producer swaps choices until they sound suitable. That process breaks when one character has several names, a crowd contains unnamed speakers, or narration slips into dialogue without a clean tag. A real casting system keeps those failures visible before thousands of lines are generated. The review time belongs inside [B][URL='https://goldmidi.com/community/threads/what-ai-narration-did-to-audiobook-production-costs.76576/']the full cost of AI audiobook production[/URL][/B] because a cheap render becomes expensive once a mistaken assignment spreads through every chapter. [HEADING=2]A character registry stops silent casting errors[/HEADING] Start with a canonical record for every speaking identity. Each record needs a stable character ID, printed name, aliases, titles, pronouns, role size, first appearance, and the approved voice. “Doctor Hale,” “the surgeon,” and “Mara” should resolve to one person when the manuscript uses all three. Unnamed speakers need rules rather than guesses. A guard who returns in chapter twelve should not become a new voice because the earlier scene called her “a woman at the gate.” Mark uncertain identities for review and preserve the surrounding line, chapter, and scene so an editor can decide with context. The registry should distinguish identity from performance mode. A character heard as a child, an adult, through a telephone, or while disguising a voice may need controlled variations. Those versions remain attached to one identity, preventing the system from merging different people or splitting one person into an accidental cast. Minor roles also need a declared policy. You can reserve unique voices for principal and recurring characters, then use a small approved pool for walk-ons. The assignment should consider whether pooled roles ever appear together, since reusing one voice inside a shared scene can make two speakers sound like one person. [HEADING=2]Voices must separate inside the same scene[/HEADING] Casting from isolated samples rewards voices that sound attractive alone. It does not reveal whether the narrator and protagonist occupy the same pitch, pace, accent, or vocal weight. Audition the actual exchanges where important characters interrupt, argue, whisper, or speak in quick succession. That review should measure [B]scene-level voice contrast[/B] rather than chase maximum variety across the whole book. Two voices can both fit their character descriptions and still blur when placed side by side. A scene map shows which pairs meet often, allowing the most important pairings to receive the clearest separation. The narrator needs a place in that map. Narrative prose often touches dialogue without a long pause, and some books move between external narration and a protagonist’s thoughts. Test those boundaries with the intended spacing and mastering chain instead of judging clean voice clips from a library page. Accent and age labels should remain supporting notes, not final decisions. The useful test is whether the listener can follow the exchange and whether the performance fits the text over several minutes. Short auditions hide repetitive cadences, exaggerated mannerisms, and emotional settings that become tiring across a full chapter. Approval clips should include a neutral passage, an emotional passage, and a crowded scene for each principal role. Save those clips beside the character record. They become comparison material when later chapters sound different or a platform replaces the underlying voice model. [HEADING=2]Version control protects the finished performance[/HEADING] A cast list that stores only display names is fragile. Library names can change, voices can be removed, and model updates can alter delivery even when the visible selection looks unchanged. Record the provider, voice identifier, model version, language, stability settings, pace instructions, and approval date. Freeze that manifest before bulk generation. Current audiobook tooling can update every line assigned to a character when its voice changes, clearing earlier audio and requiring those passages to be generated again. A late swap therefore has a measurable blast radius in credits, review time, and mastered files. The production log should connect each rendered passage to a character ID and manifest version. When an editor corrects an alias or speaker boundary, the team can regenerate only affected lines and identify chapters that require another listening pass. Without that link, a small casting correction becomes a manual search through the entire book. Model changes deserve the same control as cast changes. Re-audition the saved reference scenes before accepting a new version, then record the decision and every regenerated chapter. The finished performance stays reproducible only while its speaker map, voice settings, approvals, and rendered files remain tied to the same locked record. [/QUOTE]

Name