2026-03-22
By Vadym · Generated with Boba, curated by me
March 2026
Last week, I recorded a three-second voice clip — just me saying "hello, this is a test" — fed it into an open-source AI model, and got back a synthetic version of my voice that even I had trouble telling apart from the real thing. No cloud API. No subscription. No special hardware. Just a MacBook and a free download.
The model is called Qwen3-TTS, released by Alibaba's Qwen team under an Apache 2.0 license. It runs locally on Apple Silicon through a library called mlx-audio. Setup took about ten minutes. The result was unsettling.
Voice cloning used to be a studio operation — clean recordings, expensive software, time. Commercial tools lowered the barrier. That was already alarming.
Now it's different. Three seconds. Zero dollars. Completely offline.
Qwen3-TTS isn't alone. Fish Speech, CosyVoice2, IndexTTS-2 — there's a growing ecosystem of open-source voice cloning models. But Qwen3-TTS stands out because of how little it asks from the user: a tiny audio sample, a consumer laptop, and almost no friction.
That's the part that matters. Not that the model is impressive — it is. The real shift is that voice cloning is no longer a specialized capability. It's becoming a default one.
Credit where it's due: this technology is genuinely useful.
People with speech impairments can communicate in a voice that sounds like them. Podcasters can produce multilingual content without hiring voice actors. Indie developers can add narration without studio budgets. One blogger added audio narration to an entire site using his own cloned voice — something that would have cost real money not long ago.
I've been using it to generate voice replies in messaging apps, narrate documents, and test audio workflows. Cross-lingual cloning works too: clone a voice in English, generate speech in Japanese, and the result is surprisingly natural.
The upside is real.
Deepfake-enabled fraud cost hundreds of millions in early 2025 alone. The UN described voice cloning this month as a "global wake-up call to organised fraud."
The cases are already ugly. A finance worker at a multinational firm transferred $25 million after a video call where the CFO — and every other person on the call — were deepfake simulations. Fortune called 2026 "the year you get fooled by a deepfake."
Those were attacks using commercial tools with guardrails.
Qwen3-TTS has none.
It's Apache 2.0. Anyone can run it locally. No identity checks. No usage logging. No content filters. Just raw capability sitting on a laptop.
The old grandparent scam — a caller pretending to be a family member in distress — just got an upgrade. All an attacker needs is a few seconds of your voice from a TikTok, an Instagram story, a podcast clip, or a voicemail greeting.
There’s a scene in Terminator 2 where a machine mimics a trusted voice over the phone to get past human suspicion. At the time, that felt like science fiction. Now it feels like a useful reference point: we may not have the robot, but we do have the imitation layer — and that part is already real.
This genie isn't going back in the bottle. But you can make yourself harder to exploit.
Set up a family safe word. A verbal passphrase only your family knows. If someone calls claiming to be your kid in an emergency, ask for the word first. This is the single cleanest defense against voice-clone scams.
Treat urgency as a warning sign. Every scam leans on panic. "Don't tell anyone." "Do it now." "There's no time." The pressure is the signal.
Call back on a known number. If your "boss" asks for a wire transfer, hang up and dial their real number. Every time. No exceptions.
Treat your voice like a biometric. Your public videos, voicemail greeting, podcast appearances, and social posts are training data now. You can't undo that, but you can stop pretending it isn't true.
Stop trusting voice verification. Any institution still using voiceprint-style verification as a security layer is behind the curve. That system is already broken.
The individual tips above are good hygiene. But organizations face a different problem: they have employees who will receive spoofed calls from "the CEO," "legal," or "IT" — and those employees are trained to comply quickly.
A few practical steps that actually move the needle:
Create verbal authorization codes for high-stakes requests. Wire transfers, credential resets, access grants — any of these should require a verbal passphrase that isn't shared in writing anywhere. One word, changed quarterly. This is cheap and almost impossible to spoof if implemented consistently.
Kill voice-only approval for financial transactions. Voice calls should never be the sole authorization channel for anything involving money or access. Require a second channel — email, ticketing system, callback to a known number — every time, no exceptions.
Brief your team on the capability, not just the policy. Most employees still think deepfake voice is a sci-fi threat. A five-minute demo changes that. Show them it works. The best social engineering defense is employees who don't find it unbelievable.
Audit your public voice footprint. Executive interviews, earnings calls, conference talks, podcasts — these are all training data. You can't delete them, but you can factor them into your threat model. High-profile voices warrant higher verification thresholds.
The technology isn't going to pause while organizations catch up. The gap between capability and awareness is where most damage happens.
I built Qwen3-TTS into my daily workflow. It's one of the best open-source AI tools I've used this year.
And every time I run it, I think about how easy it would be to point it at someone else's voice. No special skill required. No barrier worth mentioning.
We're in a short window where the capability exists but public awareness hasn't caught up. That window is closing fast. The best thing we can do is talk about this plainly — before the next wave of scams teaches the lesson for us.
Because the call you get from your daughter next month might not be her.