How much audio does it take to clone a voice?

A few seconds of clean speech can be enough. In a 2023 research paper, Microsoft's VALL-E model produced personalized speech from a 3-second recording of a speaker it had never heard before ¹. Starling Bank, a UK bank, says criminals can now clone a person's voice from as little as three seconds of audio ². Those are the floor for a demonstration, not a guarantee that a clone will fool someone on the phone. For a voice clone built with a commercial tool, the vendor asks for more.

Commercial tools ask for more, and say audio quality matters most

ElevenLabs, a voice AI company, recommends about 1 to 2 minutes of clear audio for its instant voice cloning, with no reverb or background noise ³. It also says audio quality is the most important factor in instant voice cloning ⁴.

The difference between these numbers is about what a clone needs to sound good, not about what it needs to exist. A research demo can start from a few seconds. A consumer tool that wants a stable, natural result asks for more and cleaner speech. A scammer does not need the best result, only one that is good enough for a short, urgent call.

Public audio can be enough, and the speaker never has to agree

You do not need access to a private recording. Consumer Reports tested six voice cloning products, and its researchers easily created a clone from publicly available audio in four of them ⁵. The same report found that a majority of the six lacked meaningful safeguards against fraud or misuse ⁶.

Public audio includes a video, a podcast, a voice note, a conference talk or a voicemail greeting. The speaker never has to cooperate, so the amount of audio you post is the amount that can be used. For more on stored copies, see the question on whether your stored voice recordings can be used to clone you.

It depends on the noise, the use and the listener

The answer changes with the conditions:

• Clean speech beats long speech. Vendor guidance asks for audio without reverb or background noise, so a long noisy recording can be worse than a short clean one.
• A short call needs less than a long conversation. A caller who speaks for twenty seconds on a phone line has far less to get right than a fake that must hold up across a meeting.
• Listeners are not reliable judges. In a 2023 PLOS ONE study, people spotted speech deepfakes only 73% of the time ⁷. A clone that is good enough to pass a quick check does not need to be perfect.

Our sources do not measure how much audio is needed to fool a family member on a phone line, so treat the 3-second figure as the sample length used in that research, not as a rule for a real call.

What to do: assume a few seconds of your voice is enough

1. Treat every clear recording as usable. A short video or voice note that is public should be treated as a sample someone can use.
2. Keep your voicemail greeting generic. A greeting is a recording of you that anyone can call and hear.
3. Do not use your voice as the check. A familiar voice is not proof of identity, so verify requests through a separate channel. The question on how to protect yourself from voice cloning scams sets out those steps.

The short answer is that a few seconds can be enough to start, so treat any clear public recording of your voice as usable. For the wider picture, see our guide to voice cloning and deepfakes, which covers the scams, the regulators and the defenses, or the overview of voice privacy.