You commented CLONE. Here's the paper behind the reel, the official advice, and a Claude prompt that sets up your family's safe word and call-back plan in about ten minutes.
1. How three seconds becomes a voice
The paper is "Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers" (VALL-E), from Microsoft. Its claim: the model can "synthesize high-quality personalized speech with only a 3-second enrolled recording of an unseen speaker as an acoustic prompt."
It learns from a lot of speech. The researchers trained it on 60K hours of English speech from over 7000 speakers.
It treats speech like language. It turns audio into discrete codes using an audio codec, then predicts those codes the way a chatbot predicts words. In their words, it regards TTS "as a conditional language modeling task rather than continuous signal regression."
The 3-second clip is the prompt. It carries the voice. The paper says the model could "preserve the speaker's emotion and acoustic environment of the acoustic prompt."
It beat the best earlier system. Per the paper, it "significantly outperforms the state-of-the-art zero-shot TTS system in terms of speech naturalness and speaker similarity."
The caveat, because I said I'd be straight with you: this is a research model, tested on clean LibriSpeech samples of 4 to 10 seconds. The authors say it "still suffers from several issues" and that "some words may be unclear, missed, or duplicated." They also flag the risk themselves: it may be misused for "spoofing voice identification or impersonating a specific speaker." The paper does not say scammers use VALL-E. It shows why a short clip is enough to worry about.
2. The paper and the official advice
The paper: VALL-E, arXiv 2301.02111
FBI: IC3 public service announcement, 3 Dec 2024. Criminals "generate short audio clips containing a loved one's voice to impersonate a close relative in a crisis situation." Its first tip: "Create a secret word or phrase with your family to verify their identity."
FTC: Scammers use AI to enhance their family emergency schemes. A scammer needs "a short audio clip of your family member's voice", which he could get from content posted online. See also Scammers use fake emergencies to steal your money. The FTC says don't trust the voice, call back on a number you know is theirs, and call someone else in the family even if the caller says to keep it secret. (Those FTC pages don't mention a safe word. That bit is the FBI and Starling.)
Starling Bank: Safe Phrases and the campaign launch. They say criminals can clone a voice "from as little as three seconds of audio", easily captured from a video someone has uploaded online or to social media. If the caller can't give the phrase: "End the call and do not make any payments." In the UK, call 159 to reach your bank.
3. The Claude prompt
Paste this into Claude (any chat works). It asks one question at a time, so it takes about ten minutes with the family in the room. Do the safe word part together, in person.
You are helping my family set up protection against AI voice-clone scams. Work through the steps one at a time. Ask me one question at a time and wait for my answer.
1. Safe word. Ask who is in our circle. Help us choose a safe phrase that is random but easy to remember, not on any social profile, and not a pet name, birthday or address. Offer five examples for inspiration, but tell us to pick our own. Tell us to agree it in person and never type, text or email it. Remind us to change it every now and again, and straight away if we think it leaked.
2. Call-back plan. Ask me for each person's real number, one by one, and build a known-numbers list. Rule: if anyone calls asking for money or urgent help, hang up and ring back on a number from the list, or call another relative.
3. Mid-call lines. Write five short lines we can say, such as "What's our safe word?" and "I'll call you straight back." Include the red flags: urgency, secrecy, and payment by gift card, crypto or wire transfer mean stop.
4. Printable card. Finish with a one-page card: the known-numbers list, the mid-call lines, and a blank space to write the safe word by hand. Tell us to keep the card and the word apart.
Use plain English. Never ask me to tell you the actual safe word.Starling's own checklist for a phrase: simple yet random, easy to remember, different from your other passwords, shared in person. The "never type it" rule matters because the point is that it never goes digital. Change it every now and again, and immediately if you think it's no longer private.
4. If you get the call
Stop. Urgent plus "you're the only one who can help" is the FTC's red flag. Pressure is the tell.
Ask for the safe word. No word, or a wrong one: end the call. Starling: do not make any payments.
Hang up and call back on a number you already have, not the one that just rang. Caller ID can be faked (Starling calls it spoofing). This is the FBI's advice too.
Call someone else in the family, even if the caller said to keep it secret.
Never pay by gift card, crypto, wire or payment app because a voice on the phone asked you to. The FBI says not to send money to people you have only met over the phone.
Spoke to your bank? UK: call 159 to contact your bank immediately.
One more cheap fix from the FBI: if you can, limit online content of your voice and keep social accounts private. I'm not telling you to delete your reels. I'm telling you to set the safe word.
Mike
PS: Want more like this, how the tech you use every day actually works? You're already on the list. Reply and tell me what to break down next.