Voice, consent, and the rules we set for ourselves

A voice is not just a sound. It is how people recognize their parents on the phone, how colleagues know who is speaking in a meeting, how a public figure's words carry authority. It is part of someone's identity in a way that is easy to underestimate until it is misused. As we build speech models, we have had to decide what we will and will not do with that capability. This post sets out the principles we hold ourselves to, why we hold them, and where we think the hard questions still are.

Why this matters now

Only a few years ago, producing a convincing imitation of a specific person's voice required hours of clean recordings and considerable expertise. Today, research systems and commercial tools can produce a recognizable likeness from a short sample. That progress has real benefits: people who have lost their voice to illness can preserve it, creators can localize their work, and accessibility tools can sound less mechanical. But the same capability lowers the cost of harm.

The risks are not speculative. Voice-based fraud, where a caller imitates a family member in distress or an executive authorizing a payment, has been widely reported. Synthetic audio can put words in a public figure's mouth, and in a fast-moving news cycle a fabricated clip can spread well before it is debunked. Beyond these headline cases, there are quieter harms: someone's voice used in content they would never endorse, or a person discovering that recordings they shared casually have been turned into a model without their knowledge.

We are a small lab, and we do not pretend that our choices alone determine how this technology is used. But everyone building speech models makes choices about what they enable, and we would rather make ours explicitly.

Consent comes first

Our foundational rule is simple: we do not model a person's voice without their explicit, informed consent. That applies to the voices we train and to any voice-adaptation features we may offer in the future. Informed means the person understands what will be built, how it will be used, and who will have access to it. Explicit means an affirmative agreement, not a clause buried in general terms of service.

Consent also has to be revocable. If someone who contributed their voice later changes their mind, we want to be able to stop using the resulting voice and remove it from our systems. That is harder than it sounds, because a voice can be entangled with a model trained on many speakers, and we are thinking carefully about how to structure training so that withdrawal is meaningful rather than symbolic.

The same principle governs training data more broadly. It is tempting, especially for a lower-resource language like Albanian where clean speech data is scarce, to gather whatever audio can be found online. We have chosen not to build voices from scraped recordings of identifiable people. It means moving more slowly and investing more in collecting data we have a clear right to use. We think that is the right trade.

We would rather move slowly on data we have a right to use than quickly on data we don't.

Public figures and the uses we refuse

We will not clone the voices of public figures, including politicians, journalists, celebrities and other well-known people, without their permission. The fact that someone's voice is widely recorded does not make it available for anyone to reproduce. If anything, the reach of a public figure's voice makes an imitation more dangerous, because people are primed to believe it.

Beyond specific voices, there are uses we intend to refuse regardless of whose voice is involved. These include impersonating real people to deceive others, generating audio intended to pass as a genuine recording of events that did not happen, harassment, and attempts to defeat voice-based authentication systems. Where our models are offered through products, we plan to write these limits into our terms, build technical checks where they are feasible, and act when we find violations.

We are realistic about the limits of rules like these. A determined bad actor with enough technical skill has other options, and no policy can prevent every misuse. But the fact that a boundary cannot be made perfect is not a reason to leave it undrawn. Clear lines make it easier to design safeguards, to explain decisions, and to be held to account.

Provenance and watermarking

Policies govern what we do; provenance tools help everyone else tell what has been done. We are researching ways to mark synthetic audio so that it can later be identified as machine-generated. Audio watermarking embeds a signal in the waveform that is imperceptible to listeners but detectable by software. The hard part is robustness: a watermark needs to survive compression, re-recording through a speaker and microphone, speed changes and deliberate attempts to strip it, while leaving the audio quality untouched.

Watermarking is not a complete answer. It only covers audio produced by systems that choose to embed it, and detection tools are only useful if they are widely available and trusted. We see it as one layer alongside others, such as metadata standards that record where content came from, and detection research aimed at synthetic speech in general rather than any single system's output. We intend to share what we learn about what works and what does not, because provenance is only valuable if it is broadly adopted.

Thinking about the trade-offs

Some of these decisions carry real costs, and it would be dishonest to present them as free. Strict consent requirements make it harder to build voices for underrepresented languages, which is one of the things we most want to do. Restricting voice adaptation limits legitimate uses, such as someone preserving their own voice ahead of a medical procedure, or a creator dubbing their own content into another language. Watermarks that are robust enough to matter may, if poorly designed, subtly affect quality.

Our approach to these trade-offs is to start from the person whose voice is involved. Most beneficial uses of voice technology are ones where that person is present and willing: they want their own voice preserved, or they have agreed to lend it to a product. Designing around consent tends to preserve those uses while closing off the harmful ones, which usually depend on someone's voice being used without their knowledge. Where a use case falls in a grey area, we would rather take the time to understand it than default to either extreme.

We also expect to get some of this wrong. The technology is moving quickly, norms are still forming, and regulation is taking shape in different ways in different places. We work across the United States, Kosovo and the United Kingdom, and we pay attention to how expectations differ between them. We treat these principles as commitments we will revisit as we learn, not as a finished policy, and when we change them, we intend to say so openly.

What we're committing to

To put it plainly: we will not model anyone's voice without their explicit consent, and we will honour requests to withdraw it. We will not reproduce the voices of public figures without permission. We will refuse uses built on deception, impersonation or harassment. We will invest in provenance research and share what we learn. And we will be open about where these commitments constrain our work, because the constraints are part of the point.

Speech technology can make computers far more useful and accessible, and we are excited to help build it. Doing that well means taking seriously that every voice belongs to someone.