The supplier should obtain a specific synthetic-voice agreement before training or delivering the reusable model. The recommended product is a controlled service for named users and approved purposes, with separate rights for finished recordings. Open-ended sublicensing should remain outside the initial offer unless the performer authorises it and the supplier can enforce the resulting conditions.
- Establish separate permissions for scripts, source recordings, the performance, the person's identity and personal-data processing. Ownership of the master recording does not establish the whole chain. (Copyright Act 98 of 1978, ss 6, 9, 21–22; Performers' Protection Act 11 of 1967, ss 2–5.)
- Obtain digital replica consent covering training and new speech, including approved emotional variations. Tennessee protects an identifiable simulated voice and addresses both unauthorised public use and certain person-specific cloning services. (2024 Tenn. Pub. Acts, ch. 588, §§ 3–6.)
- Make downstream permissions and technical access match the performer's grant. Separate the right to generate new speech from any continuing permission to use approved clips.
- Define withdrawal, termination, retention and takedown separately. A contractual label of irrevocability cannot remove applicable personal-data rights. (Protection of Personal Information Act 4 of 2013 (POPIA), ss 11, 14 and 24.)
Establish the rights in the source material
The supplier needs a record of who wrote each script, arranged each recording and supplied any other protected material. South African law treats literary works and sound recordings separately. A paid commission can place copyright in a sound recording with the commissioner, subject to the statutory conditions and contrary agreement. The contract and production records must establish whether that commissioner is the supplier or customer. Payment does not give the purchaser every contributor's rights. (Copyright Act, ss 1, 6, 9 and 21(1)(a), (c)–(e).)
Obtain the necessary script and recording licences for copying, preparation of training datasets, provider uploads and commercial output use. Assignments and exclusive licences require signed writing under South African law. An author's moral rights in a qualifying script also require separate consideration when approving edits; section 20 does not create an equivalent moral right in every recorded voice. (Copyright Act, ss 20 and 22(2)–(5).)
Performer consent presents a different problem. For a protected performance, reproducing a recording for a purpose outside the original consent requires fresh authorisation unless an applicable exception permits it. Building a reusable voice model is a materially different purpose from delivering conventional audio. The supplier should not rely on the agreement's silence as evidence that training was authorised. (Performers' Protection Act, ss 1, 3, 5(1)(a)(iii)(bb) and 8.)
The actual wording and circumstances could support a broader permission than the facts described establish. That possibility needs document review. The commercial route should nevertheless secure express consent rather than depend on an untested interpretation or a research exception for production deployment.
South African identity protection also matters. Grütter restrained continued commercial use of a person's name that falsely suggested a professional association. Applying that reasoning to a recognisable cloned voice requires an analogy; the judgment did not decide AI voice cloning. Obtain express approval of the identity uses and representations intended here. (Grütter v Lombard [2007] SCA 2 (RSA), paras 12–14.)
The related analysis of South African AI game assets supplied to a US publisher addresses copyright ownership and cross-border grants. This transaction adds a person's continuing interest in the use of their voice.
Apply Tennessee voice rights to the deployment
The ELVIS Act voice rights extend to a sound readily identifiable with a particular individual, including a simulation. The statute addresses knowing unauthorised advertising uses and making a voice available to the public with knowledge that the use lacks authorisation. It also reaches a service whose primary purpose or function produces a particular identifiable person's voice, where the specified knowledge requirement is met. A bespoke voice service therefore needs its own permission record. (2024 Tenn. Pub. Acts, ch. 588, §§ 3–6; Tenn. Code Ann. § 47-25-1105(a).)
The law contains First Amendment-conditioned exceptions for specified expressive uses. Ordinary branded announcements and product advertising should not be planned around an assumption that they qualify as parody, criticism or news. An AI disclosure also does not supply the performer's missing consent. (2024 Tenn. Pub. Acts, ch. 588, § 10.)
The customer's Tennessee location is a relevant connection, but it does not establish Tennessee law for every act abroad. Identify the performer's residence, where recording and training occur, where the customer deploys the voice and which audiences receive it. Affiliates distributing nationally require review of the additional states involved. Contractual governing law and forum should be chosen after that map is complete.
Federal copyright supplies a separate test. Section 114(b) limits sound-recording rights over an independent fixation that imitates other sounds. That does not excuse copying source recordings for training or settle identity rights. Check the actual training copies and output against the relevant rights and exceptions. Do not promise exclusive US copyright in every generated utterance. (17 U.S.C. §§ 102, 106–107 and 114(b).)
As of 5 November 2026, the NO FAKES Act of 2026 remains a federal proposal. It should not be described as an enacted nationwide replica right. The Copyright Office's digital-replica report is a policy recommendation, while existing federal copyright and false-endorsement law continue to matter. (S. 4591 and H.R. 8915, 119th Cong.; Congressional Record, 30 September 2026, pp. S5205–S5207; 15 U.S.C. § 1125(a)(1)(A).)
Give the performer a defined synthetic-voice grant
The replacement release should attach the approved recording set, voice model, use schedule and fees. Keep the permissions readable enough for the performer to understand what new speech can be attributed to their voice.
A proposed core grant is:
The performer authorises the supplier to use the recordings identified in the recording schedule to train, test and operate the identified voice model for the customer and affiliates named in the use schedule. Generated speech may be used only for the listed announcements and advertising, through the approved channels, in the specified territories and during the agreed term. The authorisation covers only the emotional styles and modifications expressly approved in that schedule. Further training, a new customer, an additional affiliate, a new use or transfer of the model requires further written agreement. Personal-data processing remains subject to the separate privacy terms and applicable law.
The schedules must settle the following points before signature:
- Identify training, fine-tuning, testing and inference separately. Specify the provider and model instance. Exclude provider-wide training, cross-customer reuse and retraining on generated speech unless separately authorised.
- Define announcements, campaigns, products, scripts or script categories, languages and distribution channels. Decide whether the customer can generate new wording within those categories or must submit each script for approval.
- Approve emotional styles individually. Routine changes to pace or emphasis should not silently authorise distress, sexualised speech, political advocacy or apparent personal endorsements. Specify prohibited impersonation, including authentication, fraud and representations that the performer personally said unapproved words.
- List authorised customer entities and affiliates. Agencies and hosting providers should receive limited operational access. Require equivalent restrictions downstream, a register of recipients, audit evidence and the power to suspend access. An affiliate's onward sale needs its own permission.
- State the territory, generation term, output-use term, renewal process and compensation for each permission. Price any exclusivity by the defined customers, sector and duration; do not imply ownership of the performer's identity.
The requested perpetual sublicence is a commercial demand, not a legal default. A continuing licence for identified finished announcements may be easier to control than perpetual access to a transferable model. Both still require matching performer permission and lawful data use. Qualifying US copyright grants by authors can also remain subject to statutory termination despite contrary contract wording. (17 U.S.C. § 203(a), (a)(5), (b)(5).)
The customer agreement must grant no broader use than the supplier has obtained. Reserve background software and model-provider rights, identify any licensed master recordings, and describe output rights only to the extent they exist. The same upstream and downstream matching problem appears in AI supplier contracts for EU and UK enterprise sales, although this customer's applicable law requires its own assessment.
Make data permissions and security work together
Identifiable performer recordings and associated records require a POPIA assessment. Voice recognition used for personal identification can involve biometrics; an audio file is not automatically biometric information. Determine whether the model, embeddings and generated clips remain information about an identifiable person. Removing a filename is insufficient to establish de-identification. (POPIA, s 1.)
Record the lawful basis and purpose for each operation. Where biometrics are processed, assess the additional authorisation provisions, including sections 27 and 33. Give the performer clear information about training, recipients, retention and rights. A supplier deciding to reuse recordings for its own model development must account for that purpose, even if another activity is performed only on customer instructions. (POPIA, ss 8–11, 13–15, 18, 26–27 and 33.)
Use written operator terms, controlled access and a documented export route before a foreign provider receives personal information. A binding agreement can support section 72 protection if it meets the statutory standard. Transfers of special personal information to a recipient without adequate protection also require the section 57 prior-authorisation assessment and compliance with section 58. The destination's name alone does not decide either question. (POPIA, ss 20–21, 57–58 and 72.)
The technical lead should isolate the performer-specific model component, restrict downloads, log generation and test access revocation. The supplier should prevent unauthorised users from obtaining reusable samples or model files. Record who investigates a leak, disables generation, preserves necessary evidence and handles notifications. These are proposed implementations of risk-appropriate security obligations. (POPIA, ss 19–22.)
If announcements become outbound telephone calls, assess the calling rules before activation. The Federal Communications Commission treats AI-generated voices as artificial voices under the Telephone Consumer Protection Act; consent and exemption rules depend on the call and destination. The performer's release does not supply the called person's consent. The Philippine voice-agent analysis addresses a different deployment and illustrates why a channel change needs its own review. (FCC 24-17, paras 1, 4–6; 47 U.S.C. § 227(b).)
Agree the exit before the first generation
Separate expiry of the generation term, termination for breach, withdrawal of data consent and a substantiated misuse complaint. POPIA permits withdrawal of consent while preserving the lawfulness of prior processing and processing supported by the other statutory grounds. A contractual non-exclusive copyright licence has its own revocation rules. Neither event should be described as automatically determining every other permission. (POPIA, s 11(2)(b); Copyright Act, s 22(4).)
The proposed exit schedule should require:
- Immediate suspension of new generation and new sublicensing when the relevant permission ends or a defined serious-misuse trigger occurs. Notify every authorised downstream user and revoke their credentials.
- Removal of samples, caches, training copies and performer-specific model components according to a documented retention and deletion schedule. Any necessary legal archive must be restricted to that purpose. Backups need an expiry process and controls preventing restored data from re-entering production.
- A technical assessment of any shared model trained on the recordings. If the supplier cannot support the promised cessation or deletion, it should change the architecture or use another performer before launch. A deletion certificate for the source files alone does not establish that model use has ended.
- A separate rule for approved clips already distributed. Identify any permitted continuing uses, channels and end dates. Stop active campaigns and controlled distribution where permission no longer supports them; request downstream takedown and document the response. Do not promise universal recall of copies outside the parties' control.
Retention exceptions and continued output use still need an applicable legal basis. A paid invoice or a survival clause cannot settle that assessment by itself. (POPIA, ss 11 and 14.)
Assemble the release and launch records
The legal and production leads should first verify identity, capacity, existing exclusive commitments and rights in every input. They should then complete the performer grant, customer licence, permitted-use schedules, privacy information and provider terms. Require evidence that the provider supports the promised restrictions. Allocate responsibility for cleared inputs, customer advertising claims, unauthorised use, defence of covered claims and replacement of an unusable voice.
The finance lead should also check the South African outbound intellectual property (IP) licence route. Where the supplier is a South African exchange-control resident licensing its IP to the non-resident customer, the Authorised Dealer permission requires the licence and supporting fee evidence, with repatriation and tax conditions. Address that route before the operative grant. (SARB, Currency and Exchanges Manual for Authorised Dealers, 8 October 2026, B.2(A)(ii)–(v).)
Before release, test an unapproved affiliate, an excluded emotional style, credential cancellation and removal of an active campaign. Keep the approved version, permissions and test results together. If the performer declines the required grant, offer separately cleared conventional recordings, recruit a willing performer, or postpone the reusable product. The existing release supplies no demonstrated basis for the customer's requested unrestricted model rights.
Scope the voice rights review
Contact Licentium to scope a review of the performer release, customer terms and intended uses for an AI voice product. Include the recording agreement, proposed model workflow and requested affiliate and sublicensing rights.