Last verified: July 19, 2026. This article distinguishes Google's Lyria 3 models from the other music providers available in a multi-model studio.
The short answer
Google states that audio generated by its official Lyria 3 models includes an imperceptible SynthID audio watermark. That statement applies to output from lyria-3-clip-preview and lyria-3-pro-preview; it should not be generalized to Suno, MiniMax, or another provider.
On lyria-3.pro, check the model and provider label shown in the generator. A track produced by a non-Google provider is not automatically a SynthID-marked Lyria 3 track.
Create a track and review the provider before generating
Which Lyria 3 outputs include SynthID?
Google's current Gemini API documentation describes two preview models:
| Official model | Typical output | SynthID |
|---|---|---|
lyria-3-clip-preview | 30-second MP3 clip | Included |
lyria-3-pro-preview | Full song lasting a couple of minutes; MP3 or requested WAV | Included |
| A third-party Suno or MiniMax model | Depends on that provider | Do not assume SynthID |
Both official Lyria models generate 44.1 kHz stereo audio and support text or image input. Google documents SynthID as a model limitation/feature for all generated Lyria 3 audio.
Primary sources: Google's Lyria 3 generation guide and Gemini API pricing.
What SynthID does
SynthID is a provenance signal embedded into generated content. For audio, the signal is designed to be imperceptible during normal listening. It differs from ordinary file metadata:
| Provenance method | Stored where | Can normal export strip it? |
|---|---|---|
| Title/artist metadata | File tags | Often |
| C2PA content credentials | Signed metadata and manifest | Can be separated from a copied asset |
| SynthID audio watermark | Audio signal created by the model | Designed to survive common transformations |
Watermark detection is evidence about the generation source; it is not proof of copyright ownership, a commercial-use license, or factual accuracy.
What creators should record
Keep a small provenance record for every important export:
- Provider and exact model ID.
- Generation date and task ID.
- Prompt and any uploaded source assets.
- The provider terms that applied on that date.
- The original downloaded file before later edits.
This record remains useful even when a distribution platform strips metadata. It also prevents a common mistake: claiming that every AI-generated song has Google's watermark when only official Lyria output does.
Can you remove or verify SynthID?
Google describes SynthID as robust to common modifications, but no watermark should be treated as an indestructible rights-management system. Heavy editing, analog re-recording, or short excerpts may make any signal harder to detect. Do not promise that a watermark can “always” be recovered after arbitrary processing.
Google has also exposed SynthID verification in supported Gemini experiences. Availability can vary by product and account, so use Google's current product instructions rather than an old screenshot or third-party detector.
SynthID does not grant usage rights
Three questions are separate:
- Provenance: Was this audio generated by an official Lyria model?
- Usage rights: What do the provider and platform terms permit?
- Input rights: Did the prompt, lyrics, voice, sample, or uploaded image infringe someone else's rights?
A SynthID signal only helps with the first question. Review the relevant provider terms for the other two, especially for advertising, client delivery, music distribution, or voice likeness.
A practical workflow
Use a fast clip to validate the direction, preserve its prompt, then create a full version only when the concept works. In a multi-model workflow, compare outputs by listening quality and provider terms—not by assuming every option shares Google's provenance system.
Try one generation with a clearly labeled model or compare the current plans.