Ten seconds of audio can now clone a voice — which means your voice has stopped being proof of who you are

Ten seconds of recorded speech is now enough to clone a human voice with a realism that was science fiction a couple of years ago, according to El Destape, citing WIRED. The company behind the jump is ElevenLabs, which specialises in hyper-realistic audio synthesis. The figure is striking, but the real story is what it makes obsolete: the voice on the other end of a phone call has stopped working as evidence of identity, and most of the institutions that still treat it as one — banks, support lines, families taking an urgent call — have not updated their assumptions.

The technical details matter. Per the report, the system does not merely copy a tone from a short sample: it keeps the vocal identity stable across long texts, and lets the user steer expressiveness through tags ranging from calm to enthusiastic. Language coverage went from 70 to more than 90 — meaningful in markets where synthetic speech until recently sounded robotic or simply did not exist outside English. The company's stated uses are legitimate and, in several cases, socially valuable: dubbing, accessibility, content production, audiobooks, corporate training and advertising. It has also become a core component of conversational agents and phone support centres, where a low-latency synthetic voice — one that replies fast enough to feel like a live conversation — can handle queries without the caller noticing the difference.

That last sentence is where the problem lives. A voice indistinguishable from a human's, available on demand, is the same capability whether it answers a billing question or impersonates your daughter. Ten seconds of audio is not a barrier: as the article notes, it can be lifted from a public video or a forwarded voice note. The result is a straight path to phone scams, identity impersonation and audio deepfakes — synthetic recordings passed off as real — that are increasingly hard to tell from genuine ones. The piece identifies the core of the concern precisely, and we agree with the framing: innovation is moving faster than the consent and verification mechanisms that should accompany it.

This fits a pattern we have argued repeatedly. The nearest-term AI danger is not a distant superintelligence; it is the industrialisation of fraud. What changed is not that voice cloning became possible — it was possible before — but that it became cheap, fast and self-service. Cloning once required a long clean sample and expertise; now it requires a WhatsApp audio and a subscription. Cost collapse is the attack vector. And refusing to build the technology does not make it disappear: open-weight audio models, the kind anyone can download and run on their own hardware, are climbing the same curve. The capability is out. The question is what sits underneath it.

Our reading: when generating is effectively free, both the value and the bottleneck move to verifying. The missing piece here is infrastructure, not restraint — out-of-band verification as the default for any money movement triggered by a call, signed provenance travelling with audio files, consent registries for voice likeness, and liability placed where the money actually moves rather than on the 70-year-old asked to recognise a voice under pressure. The article's advice is sound and worth repeating — distrust urgent calls demanding money or sensitive data on the strength of a familiar voice, and always verify through a second channel — but it is a stopgap. A security model that depends on every citizen being permanently suspicious is not a security model; it is an unpaid tax on the most vulnerable.

The long view is still worth defending, and it is not naive. The same system that enables the scam gives speech back to people who have lost theirs to illness, makes a book readable to someone who cannot read it, and carries knowledge across more than 90 languages instead of a handful. That is real abundance, and it arrives this decade, not in some speculative future. But the transition bill is being paid right now, by people who never agreed to pay it. Ten seconds is the new attack surface. The answer is not to stop speaking in public — it is to stop treating a voice as a password, and to build the verification rails before the next wave of fraud, not after it.