Voice Cloning
Also known as: voice synthesis, AI voice cloning, voice deepfake, synthetic voice
Voice cloning uses AI to analyze recordings of a real person's voice and produce a model that can generate new speech in that voice from any typed text.
High-quality modern voice cloning requires only a few seconds to a few minutes of audio samples. The technology is used legitimately for accessibility (restoring a person's voice after illness), entertainment (voiceover production), and video game character voices, but also creates significant risks of fraud, non-consensual impersonation, and audio deepfakes.
Voice Cloning, explained at three levels
At home
Voice cloning is when an AI listens to recordings of someone's voice and learns to copy it. Once the AI has learned the voice, you can type any sentence and it will speak those words in that person's voice. This is why you may have heard stories of scammers calling people and pretending to be a family member in distress: they use a cloned voice to make the impersonation convincing.
At work
Voice cloning has legitimate uses in professional contexts: a content creator might clone their own voice to produce audiobook narrations without recording every word; a TV broadcaster who loses their voice to illness might use a clone to continue working; video game studios clone voice actors to produce character dialogue variations efficiently. In all legitimate cases, the person whose voice is cloned should have given informed consent.
At home
Voice cloning models learn to represent the distinctive acoustic features of a person's voice (their pitch, rhythm, breathiness, accent, and vocal timbre) from audio samples. The model then applies these characteristics to any input text, producing speech that sounds like that person saying those words. Modern models can achieve convincing results from as little as three seconds of audio, making the barrier to cloning very low.
At work
The fraud risk from voice cloning is real and growing. 'Virtual kidnapping' scams use cloned voices of alleged family members to demand ransom payments. Business email compromise attacks have evolved to include cloned-voice phone calls impersonating executives authorizing wire transfers. In early 2024, a finance employee at the engineering firm Arup in Hong Kong was tricked into transferring about US$25 million after a video call appeared to show the company's CFO and other colleagues, all of them AI-generated deepfakes, instructing the transfer.
At home
Detection of AI-generated voice is increasingly difficult. Human listeners often cannot distinguish high-quality cloned voices from real ones, particularly over phone calls where audio quality is already compressed. AI voice detection tools exist but have limited accuracy, especially against voices they have not seen in training. Watermarking of synthetic audio (embedding inaudible markers in AI-generated speech) is a developing countermeasure that some providers implement.
At work
From a practitioner standpoint, organizations should consider voice cloning in their threat model for fraud and social engineering. Caller verification procedures (code words, callback to known numbers, transaction approval through a separate authenticated channel) are the most reliable defence against voice impersonation attacks. For legitimate use, consent and provenance are the key requirements: the person whose voice is used should have provided explicit informed consent, and AI-generated audio should be disclosed where relevant.
Real-world examples
At home
After losing his voice to throat cancer, retired teacher Carlos worked with a specialist company to clone his voice from home recordings before surgery. The resulting voice model now reads his digital notes and messages aloud in his own voice, allowing him to continue communicating in a voice his family recognizes.
At work
An audiobook publisher used a voice cloning service (with the narrator's written consent and compensation agreement) to produce a 200-chapter series from a single narrator's eight-hour reference recording, dramatically reducing studio time while maintaining the narrator's distinctive performance style.
Common confusions
- Voice cloning is not the same as text-to-speech. Generic text-to-speech produces a synthetic voice that belongs to no specific person. Voice cloning produces a synthetic voice that is specifically modelled on and sounds like a named individual. The distinction matters enormously for consent and fraud risk.
- A phone call that sounds like a family member is not proof that the family member is on the line. As voice cloning becomes more accessible, caller ID and the sound of someone's voice are no longer reliable authentication factors. Sensitive decisions triggered by a phone call (particularly financial transactions or sharing personal information) should use a separate, out-of-band verification step.
- Voice cloning without consent is a harm even without fraud. Cloning someone's voice to produce audio they never said (even for entertainment) can damage reputation, cause emotional distress, and in some jurisdictions is addressed by emerging AI and identity laws. Several US states have passed legislation specifically targeting non-consensual voice and likeness cloning.
For analysts
In BA artefacts, voice cloning risks surface when designing authentication and verification workflows. Voice authentication (using voice as a biometric factor) should be assessed against the risk that voice can be cloned. For systems where voice authentication is used, specify that it should not be the sole authentication factor for high-value transactions. When building legitimate voice cloning features into products, specify consent workflow requirements: whose voice is being cloned, how consent is obtained, how it is recorded, and how the cloned model is protected against unauthorized use.
History & origin
Text-to-speech synthesis has existed since the 1960s, but voices were robotic and clearly artificial. Neural text-to-speech (2016-2020) dramatically improved naturalness, making synthetic voices indistinguishable from humans in many contexts. Voice cloning (creating a personalized voice model from a specific person's recordings) became practical at scale when tools like ElevenLabs (2022) demonstrated that just seconds of audio could produce a convincing clone. By 2024, real-time voice cloning (transforming a speaker's voice into another person's voice mid-conversation) had become available, raising new fraud risks.