Reply to thread

Message: [QUOTE="Bombastus, post: 91171, member: 2178"] SAG-AFTRA says its audiobook contracts protect narrators against digital replicas created or used without informed consent and bargaining. That protection matters because audiobook agreements vary among publishers and producers, leaving the signed language to carry much of the practical detail. A cloning deal involves at least three assets. They are the source recordings, the voice model, and every finished audio file generated from that model. Permission to record ordinary narration does not quietly become permission to train a system, synthesize new speech, or hand files to another company. Those rights need their own bargain even when [B][URL='https://goldmidi.com/community/threads/what-ai-narration-did-to-audiobook-production-costs.76576/']the production economics of authorized voice replicas[/URL][/B] make synthetic narration attractive. A lower production bill says nothing about who can approve a later title, translation, advertisement, or character performance. [HEADING=2]Creation consent does not authorize every use[/HEADING] Consent to create a replica and consent to use it should appear as separate grants. The contract should identify the performer, client, model operator, approved technology, and exact recordings that may enter training. It should also say whether the model can incorporate another voice or any blended dataset. A protective default is simple. No unlisted material enters the model, and no model or recording goes to a third party without the performer’s written approval. That restriction should follow the assets through contractors, affiliates, acquisitions, and any replacement service provider. The use grant then needs a tight boundary. [B]Title-specific replica permission[/B] can name the book, edition, language, territory, distribution channels, term, characters, and permitted promotional excerpts. A sequel, translation, adaptation, game, sample pack, or unrelated narration becomes a new use rather than a creative extension of the first deal. The content boundary deserves equal precision. Political endorsements, sexual material, defamatory passages, and uses that imitate an unscripted personal statement can require separate approval. A clause covering all media now known or later developed defeats that control. Approval also needs a defined moment. The performer might review a representative sample before full generation, then approve the mastered performance before delivery. The agreement should set correction rounds, response times, and a process for disputed pronunciations without treating silence as permanent consent. [HEADING=2]Control depends on custody and deletion[/HEADING] A promise against misuse is weak when nobody knows where the model lives. The agreement should assign custody of raw sessions, cleaned training clips, model checkpoints, generated stems, and final masters. It should name every vendor allowed to store or process them. Security terms need operational detail. Limit access to named roles, require current encryption for stored and transferred files, log downloads, and set a deadline for breach notification. The client should remain responsible when an approved vendor exposes or reuses the material. Termination requires two different decisions. One concerns future generation, which can stop immediately, while the other concerns audiobooks already approved and released. The contract should state whether existing editions may remain on sale, for how long, and under what continuing payment and credit terms. Deletion language should reach backups, vendor copies, derived datasets, and superseded model versions. A deletion certificate is more useful when it lists each system searched and each exception retained for legal or accounting reasons. If a model cannot be separated from a shared system, the agreement should address that limit before any recording begins. Retail metadata belongs in the contract too. Industry guidance distinguishes a generic AI voice from an authorized replica built from licensed samples of a specific person. Assigning responsibility for the authorized-replica label, narrator credit, and performer-name spelling prevents a licensed performance from appearing anonymous or falsely human-recorded. The performer can receive a metadata proof before delivery, while the publisher owns corrections sent to distributors. That small workflow matters when retailers display narrator identity differently or lose a disclosure during ingestion. [HEADING=2]Compensation must follow each approved use[/HEADING] One payment rarely accounts for the entire rights chain. The deal can separate session time, model creation, each approved title, pickups, performer review, translations, promotional clips, and later reuse. Clear triggers prevent a broad buyout from swallowing work that did not exist when the fee was negotiated. Revenue-based compensation needs its own definitions. The contract should identify gross or net receipts, permitted deductions, statement frequency, payment deadlines, audit access, and treatment of subscription listening. Flat fees still need due dates and an agreed price for extra correction rounds. Unauthorized use also needs a practical response. The agreement can require generation to stop, affected files to come down, copies to be deleted, and outstanding compensation to be paid while a dispute proceeds. It should allocate indemnity and enforcement costs instead of assuming an ordinary breach clause can repair an endlessly reproducible voice model. The document should also identify governing law, dispute procedure, and the provisions that survive termination. Union terms, publicity rights, privacy rules, and platform conditions may add obligations, so a client cannot treat the cloning clause as the whole agreement. [/QUOTE]

Name