Buyer guide · Text to speech
Before you monetize AI voice, check these seven things.
The question “can I use this TTS commercially?” is rarely answered by one license badge. A production voice stack can involve a model, code, voice assets, phonemizer, hosted API, cloned speaker and platform-specific content rules.
1. Separate the model license from the service terms
An open-weight model and a hosted service using that model can have different rules. If you download and run a model locally, the model and software licenses matter. If you call a vendor API, the vendor's terms, plan entitlements and acceptable-use rules also matter.
Record the exact artifact you use: model name/version, repository, service plan and access date.
2. Verify commercial rights on the plan you actually use
Do not assume the free plan has the same commercial permissions as a paid plan. Some providers explicitly attach commercial use to particular tiers. For example, ElevenLabs' September 2026 pricing page lists a commercial license in Starter, while the free tier is presented separately. Verify the current terms before monetized publication rather than relying on an old tutorial.
3. Check the voice, not only the engine
A TTS engine may permit commercial use while a specific voice has separate restrictions. Cloned voices raise additional consent and publicity concerns. Ask:
- Is this a stock/synthetic voice supplied for this purpose?
- If cloned, did the person whose voice is represented give valid permission?
- Does the service restrict impersonation, political use, deceptive use or sensitive categories?
- Can the same voice be used in ads or only ordinary content?
4. Audit the dependency chain for local TTS
Local inference can involve multiple licenses. Keep an inventory:
model weights → inference code → phonemizer → audio libraries → wrapper/API → packaged application
Kokoro-82M, for example, is currently published with an Apache 2.0 model license. Some implementations may use other components with different licenses. The practical question is not simply “is the model Apache?” but “what exactly am I distributing or using in production?”
5. Distinguish generated audio from distributing software
Software licenses often govern copying, modifying or distributing software itself. The output generated by that software can be treated differently. Never infer output ownership from the software license alone; read the relevant model/service terms. This distinction becomes especially important when you ship a desktop app, Docker image or bundled inference server rather than merely publishing an audio file.
6. Check platform rules for synthetic media
Your TTS provider is only one layer. The platform where you publish may have rules about synthetic or manipulated media, impersonation, deceptive content, political content, monetization or disclosure. A voice clip can be permitted by the model license yet still violate the destination platform's rules.
7. Save evidence of the terms you relied on
Terms change. For every production provider, keep a small record with:
- terms/pricing URL;
- date checked;
- plan name;
- commercial-use statement;
- voice/source;
- dependencies for self-hosted stacks;
- material restrictions;
- review date.
This is useful operational hygiene even if no dispute ever occurs.
A pre-publication checklist
| Question | Pass condition |
|---|---|
| Do I have commercial rights for this plan/use? | Current terms or license supports the use |
| Is the voice authorized? | Stock/synthetic voice or documented consent |
| Are local dependencies understood? | License inventory exists |
| Am I impersonating a real person? | No, unless clearly lawful/authorized and platform-compliant |
| Does the publishing platform permit it? | Relevant synthetic-media rules checked |
| Can I prove what terms I relied on? | Dated source record saved |
| Does the content itself make deceptive claims? | No |
Sources used for the examples
- ElevenLabs pricing — accessed September 8, 2026.
- Kokoro-82M official model page — accessed September 8, 2026.
Bottom line
Commercial TTS is manageable when the stack is documented. Treat license and consent checks like technical dependencies: explicit, versioned and reviewed. The worst workflow is “someone on Reddit said it was fine, so we shipped it.”