STUDY GUIDEMusic
Voice models and performers' rights
Software can now learn a voice from recordings and make it sing or speak new lines. Who owns that voice, and who gets paid, is being settled in contracts and codes.
Start here
- 1
What a voice model is
A voice model is software trained on recordings of a person until it can produce new speech or singing that sounds like them. The more clean recordings it has, the more convincing the result. Studio sessions are an ideal source.
- 2
Why session musicians are at the centre
Session players and singers are usually paid a fixed fee per session, with no royalties. If a studio can generate new performances from old sessions, it may need to book fewer of them. That is why contract wording matters so much to them.
- 3
Contracts came first, rules second
Studios added clauses granting training rights before any public rules existed, and performers often signed them as part of standard forms. Codes of practice now being drafted try to separate consent and payment for each new use.
- 4
Labelling is a separate question
Even where consent is given, listeners may not know whether a vocal was sung or generated. Some proposals require generated vocals to be labelled, which some voice-tool firms say they would accept.
Timeline
2023
Clauses appear
About 9% of reviewed Haskara session contracts include voice-model training clauses.
2025
Clauses spread
The share of reviewed contracts with such clauses reaches 41%.
2026
Most contracts
58% of contracts reviewed so far this year include the clauses.
Autumn 2026
Stoppage
About 1,200 Guild members refuse new sessions in protest.
Autumn 2026
Draft code
The Federal Office for Culture publishes a draft code requiring per-use consent and payment.
Glossary
- Software that can generate new speech or singing in a particular person's voice.
- A performer hired for a fee to play or sing on other artists' recordings.
- Permission to use recordings to teach a model, separate from permission to sell the recordings.
- Permission given separately for each new use, rather than once for all future uses.
- Marking generated audio so listeners know it was not performed live.
Where
From 320 contracts shared with Soundboard. Not a random sample.
| Label | Value (% of contracts reviewed) |
|---|---|
| 2023 | 9 |
| 2024 | 22 |
| 2025 | 41 |
| 2026 | 58 |
Who is who
group
Union representing session players and singers in the Kesh Federation.
institution
Body representing the studios that issue session contracts.
institution
Kesh federal office that has drafted a code of practice on voice models.
group
Firms, several based in Brennmark, that sell voice software to studios.
Common questions
Can a studio already use my old recordings to train a model?
That depends on what your contract says. Many recent standard forms grant the right; older contracts may not mention it at all.
Is a code of practice the same as a law?
No. A code is voluntary unless a law, contract or funding condition makes it binding, which is the main point of disagreement in Haskara.
Why would labelling matter to listeners?
It tells people whether a voice they hear was performed by a person or generated, which some see as a basic matter of honesty to audiences.
Live stories on this topic
Further reading from the sources
- What the 'synthetic derivatives' clause actually saysSoundboard · Specialist
- Voice-model clauses now in most session contracts we reviewedSoundboard · Specialist
- Federal Office for Culture publishes draft code of practice on voice modelsKesh Federal Chancellery · Primary / official
- A code that pays per use would put a price on a voiceNorthgate Policy Institute · Analysis / commentary
Demo outlets are fictional; their links do not resolve.
Study guides are written and reviewed by editors, then kept current as linked stories update. How we write them