Genie 4 can make characters speak, but the voice-training question remains

Google DeepMind’s Genie 4 can generate real-time worlds with sound and characters that respond in dialogue. A report says Google has not explained whose voices trained the model, putting consent and credit squarely in the conversation about its new capability.

Google DeepMind Watch analysis

What happened

Genie 4 is a world model that builds a scene from a prompt and continues generating it as users move through the world. Google says its new version generates audio natively, including ambient sound and interactive character dialogue. The report says access is currently limited to trusted testers, with wider access planned through Google AI Ultra and selected development partners.

The question raised by Voice Over Herald is specific: Google’s announcement describes the generated speech, but the report says it does not identify the voices or speech data used to train the model. It also notes that Google says Genie 4’s generated audio and video will include SynthID watermarking. That may help identify generated media, but it does not answer how training voices were sourced.

The article also discusses Dessa, a speech-to-speech model from Block that can listen and respond without first turning speech into text. Block says it was trained on human conversation, but, according to the report, has not specified whose conversations or how they were collected. These are separate systems and announcements, though both make voice provenance a timely question.

Why it matters

When a model can invent dialogue for characters on demand, voice is not just another output setting. It can raise questions about consent, credit and compensation for the people whose performances or conversations may have contributed to training. Those questions cannot be settled from a missing disclosure alone: the report’s account does not establish that either company used unauthorised voices.

There is also a useful distinction between identifying synthetic output and explaining its origins. A watermark may help someone recognise generated audio; it does not tell them whose material shaped the model. That gap matters to performers, developers and audiences alike.

Our read

Voice Over Herald makes a pointed case for transparency, not proof of misuse. That is the right distinction to keep in view. Companies showing off speech-generating systems should explain the provenance of the data that taught them to speak, rather than leave performers and users to infer it from a polished demo. The small but important catch: a lack of public detail is not evidence of wrongdoing.

What to watch

• Whether Google or Block publishes more detail about its speech-training data and consent practices.
• What access to Genie 4 reveals about the quality and limits of generated dialogue.
• Whether watermarking makes generated audio easier to identify in practice.

Discussion spark: For voice-generating AI, should companies have to disclose training-data sources and consent practices before a public launch, or is a clear explanation after release enough?

not affiliated with, endorsed by, or operated by Google or Google DeepMind