For most of the history of the telephone, recognising someone's voice was a reasonable way to know who you were talking to. It was never rigorous, but it worked well enough to become embedded in how families, banks and businesses handle the phone. That assumption has now failed, and it has failed quickly.
Speech synthesis has reached the point where a short sample of someone speaking is enough to produce new speech in a voice that sounds like theirs. The same capability underlies genuinely valuable applications and a category of fraud that is hard to defend against with the habits most people have. This article describes how the technology works conceptually, what it is legitimately used for, how the fraud landscape looks in general terms, and what can reasonably be done about it.
How Voice Cloning Works, Conceptually
The underlying idea is that the identity of a voice can be separated from the words being said. A speaker embedding is a compact numerical summary of what makes a voice distinctive: broad characteristics of the vocal tract, pitch range, resonance and delivery. Models trained on large collections of speech learn to place recordings of the same person close together and recordings of different people apart. That family of representations was developed largely for speaker verification, deciding whether two recordings come from the same person.
Modern synthesis reuses the idea in reverse. Instead of training a separate model per voice, one general model is conditioned on a reference sample and generates speech carrying that speaker's characteristics. This is often called zero-shot adaptation: no training run is required for a new voice, only a reference. Few-shot approaches add a little adaptation to improve similarity.
Three shifts turned this from a laboratory curiosity into a security problem:
- Sample requirements collapsed. Earlier systems needed substantial studio recordings from a cooperative speaker. Current approaches work from short, ordinary audio.
- Quality rose in the difficult direction. The remaining gap between synthetic and real speech lies mostly in prosody and long-range naturalness, exactly the qualities a telephone channel degrades anyway.
- Access broadened. Capable synthesis is no longer confined to organisations with research teams.
There is an uncomfortable symmetry here: the representations that make voice useful as a biometric are close to those that make it easy to imitate.
The Legitimate Uses Are Real
It would be a mistake to treat synthetic voice as inherently illegitimate.
- Voice banking and message banking. People facing conditions that will take away their ability to speak, such as motor neurone disease, can record samples in advance so that a communication device later speaks in something close to their own voice rather than a generic one. For them and their families this is not a convenience feature.
- Accessibility more broadly. Screen readers, audio versions of written material and communication aids benefit from natural synthetic speech.
- Brand voices. An organisation may want one consistent voice across its calls and applications, built with the informed consent and compensation of the voice artist.
- Localisation and dubbing. Carrying a performance across languages while keeping vocal character has obvious value in a country with as many languages as India.
What separates these from misuse is not the technology. It is consent, disclosure and purpose.
The Fraud Landscape
- Distress impersonation. A call or voice note appears to come from a family member in urgent trouble, asking for money immediately. The emotional pressure is the mechanism; the familiar voice removes the pause in which someone would normally verify.
- Business impersonation. A voice resembling a senior executive or a vendor contact instructs a change of payment details or an urgent transfer, often combined with secrecy and time pressure.
- Voice phishing at scale. Synthetic speech lowers the cost of running convincing scam calls in the local language rather than in stilted, obviously foreign speech.
- Attacks on voice authentication. Any process that accepts a voice as proof of identity becomes a target once imitation is cheap.
The most important consequence is simple to state and hard to internalise: hearing a familiar voice is no longer evidence of who is on the line. Every process that quietly relies on that assumption needs revisiting.
Why This Matters Acutely in India
India has an unusually phone-centred communication culture. Voice notes are a default medium for family and business conversation, so recorded samples of most people's voices already exist in places they do not control. Phone-based fraud is widespread and well organised, and there have been extensive public warnings about impersonation scams, including schemes in which callers pose as police or officials to intimidate victims into transferring money.
Two further factors sharpen the risk. Digital payments are fast and largely irreversible, which compresses the window in which a victim can reconsider. And multilingual synthesis means a scam call can arrive in the recipient's own language and register, removing an informal cue that previously made fraudulent calls feel wrong.
Defences That Actually Help
The useful defences are procedural rather than technical, because they do not depend on being able to hear the difference.
For individuals and families
- Call back on a number you already have. Not the number that called, and not a number the caller provides. This single habit defeats most impersonation attempts.
- Agree a family safe word. A word or question known only within the family, used when someone calls in an emergency asking for money, and not discoverable online.
- Treat urgency as the warning sign. Pressure to act immediately, to stay on the line and to keep the matter secret is the common structure of these calls, whatever voice is used.
For businesses
- Never treat voice as sufficient authentication. At most it is one weak signal alongside independent factors.
- Put process controls on money movement. Changes to bank details, unusual transfers and urgent payment instructions should require verification through an established channel and, above a threshold you set, a second approver. The control should be mandatory regardless of who appears to be asking, because "the instruction came from someone senior" is precisely the pretext being used.
- Remove the ability to override process by seniority. This fraud depends on staff feeling unable to pause an urgent request from a superior. Say explicitly that following the verification step is never a career risk.
- Train on the pattern, not the artefact. Staff will not reliably hear that a voice is synthetic. They can reliably notice that a request arrived out of band, was urgent, and bypassed normal steps.
- Review anything that authenticates by voice, including internal helpdesks that reset credentials after a phone conversation.
- Have a reporting route. Staff who suspect an impersonation attempt should know exactly whom to tell, quickly and without embarrassment.
Detection, Provenance and Their Limits
A serious research community works on distinguishing genuine speech from synthesised or replayed speech. The ASVspoof challenge series, running since 2015 across several editions, has been the main public benchmark, covering both synthetic speech and replay attacks. The progress is real, and so are the limits:
- Generalisation to unseen attacks. Detectors perform well against synthesis methods represented in their training data and considerably worse against methods they have not encountered. New approaches appear continuously, making this a persistent rather than a solvable gap.
- Channel degradation. Telephone audio is narrowband and heavily compressed. Many of the fine spectral cues that distinguish synthetic speech are exactly what those codecs discard, so performance on clean audio does not transfer to a phone call.
- Adversarial pressure. Detection is an adversarial setting, and systems known to be in use attract effort to defeat them.
- Operating points. At high volume, even a low false-positive rate means treating large numbers of genuine callers as fake. That has to be designed for before detection drives decisions.
Provenance and watermarking approach the problem from the other side: rather than detecting fakes, mark content at creation. Imperceptible audio watermarks can be embedded by synthesis systems, and content credential standards attach signed metadata describing how a piece of media was produced.
Responsible providers should support these, but they are not a complete answer. Watermarks must survive compression, re-recording and editing to be useful over a phone channel, metadata is easily stripped, and the approach depends on voluntary adoption by the systems generating the audio, which is precisely what a bad actor will not do. Absence of a watermark proves nothing.
Consent, Disclosure and Rights
Consent for the voice itself. Cloning a person's voice should require their informed, specific and revocable permission, with clarity about which uses are covered, for how long, and what compensation applies. A voice artist who recorded for one campaign has not agreed to an indefinite synthetic version of themselves, and consent buried in a general contract clause is weak consent.
Disclosure to the listener. Where a synthetic voice speaks to a person, the widely accepted practice is to say so plainly, early and in the listener's own language, and not to undermine that with a design encouraging them to believe otherwise. Related transparency questions are covered in our post on call transcripts and India's data protection law.
Likeness and personality rights. Indian courts have in recent years recognised that well-known individuals hold rights over the commercial use of their name, image and voice, and have granted relief against unauthorised exploitation, including imitations produced with modern tools. How far such protection extends, particularly to people who are not public figures, continues to develop, so this is a general direction rather than a settled position.
Regulatory direction. Across several jurisdictions the trend is towards transparency obligations for synthetic media, restrictions on artificial voices in unsolicited calls, and platform duties to label or act on manipulated content. In India, authorities have issued advisories and proposed measures concerning synthetically generated content and intermediary obligations. Requirements are moving, so anyone deploying synthetic voice should check the current position rather than rely on a summary.
What Responsible Providers Should Do
- Verify the right to a voice. Require evidence of consent before a voice is cloned, and make revocation a real, workable process.
- Disclose automation by default. Treat disclosure as a standing property of the product, not a setting a customer can quietly disable.
- Restrict by design. Limit creation of voices resembling identifiable individuals without authorisation, and control who may use a given voice.
- Support provenance, adopting watermarking and content credentials where practical while being honest that they are partial.
- Log and be able to answer. Keep records sufficient to investigate a complaint that a voice was misused, within data protection constraints.
- Decline some work. Applications whose purpose is to make a listener believe they are speaking to a specific person who is not there should be refused.
The broader task is cultural. The phone needs a new default assumption, in the way email eventually acquired one: a voice establishes nothing on its own, and anything consequential deserves an independent check. A family safe word and a mandatory verification step for payment changes will prevent more harm than any classifier.
At Cirio, our view is that a synthetic voice should be introduced honestly, used only with the consent of the person it belongs to, and built so that it never becomes a means of pretending to be someone else.
Frequently Asked Questions
How much audio is needed to clone a voice?
Modern zero-shot synthesis systems can produce a recognisable imitation from a short reference sample, often a matter of seconds, though longer and cleaner samples generally yield closer results. The practical implication is that publicly available audio such as voice notes, interviews or social media clips can be sufficient.
Is voice still safe to use as an authentication factor?
Voice alone should not be treated as proof of identity. Because synthesis and replay attacks can imitate a speaker convincingly, voice should at most be one signal among several, supported by independent checks such as calling back on a number already on record.
Can audio deepfakes be detected reliably?
Detection research is active and has produced systems that perform well on known attack types, but generalisation to unseen synthesis methods remains a core difficulty and accuracy tends to fall on compressed, narrowband telephone audio. Detection is a useful signal rather than a guarantee.
Should a business disclose that a caller is speaking to an AI voice?
Disclosure is widely regarded as good practice and is increasingly reflected in regulatory expectations in several jurisdictions. Telling callers plainly that they are speaking to an automated assistant builds trust and avoids the impression that a synthetic voice is being passed off as a person.