AI Transcription Just Got Private. Two New Models Run on Your Phone

Your conversations stay on your device. No cloud uploads. No data centers. No hackers lurking.

That’s the promise behind Mistral AI’s latest transcription tools, announced Wednesday. The French developer built two new models specifically designed to run locally. So your sensitive recordings never leave your laptop or phone.

Why does this matter? Think about your last doctor’s appointment or lawyer consultation. Would you want that audio file sitting in some company’s cloud storage? Probably not.

Speed Meets Privacy

Mistral AI launched two models with different strengths. Voxtral Mini Transcribe 2 handles recorded audio. Meanwhile, Voxtral Realtime works like live closed captioning.

Pierre Stock, Mistral’s VP of science operations, calls the Mini model “super, super small.” That tiny size lets it squeeze onto devices with limited processing power. Your phone. Your laptop. Even a smartwatch.

But size isn’t the only advantage. Running locally means zero network latency. No waiting for files to upload, process in distant servers, and download back to you.

“What you want is the transcription to happen super, super close to you,” Stock explained. “And the closest we can find to you is any edge device.”

Real-Time Transcription That Actually Keeps Up

Remember watching live TV with closed captions lagging three seconds behind? Frustrating, right?

Voxtral Realtime solves that problem. The model generates text with less than 200 milliseconds of latency. That’s about as fast as you can read the words yourself.

I tested both models during my interview with Stock. Voxtral Realtime captured my speech quickly and accurately. It handled English with some Spanish mixed in without breaking a sweat. Plus, it supports 13 languages right now.

However, I did notice something amusing. The transcription model got its own name wrong. It transcribed “Mistral AI” as “Mr. Lay Eye” and “Voxtral” as “VoxTroll.” Not exactly confidence-inspiring for a transcription tool.

Stock acknowledged this quirk. But he pointed out users can customize the models. You can teach them specific terminology, proper names, or industry jargon. So if you’re transcribing medical or legal conversations regularly, you can fine-tune accuracy.

Small Doesn’t Mean Weak

Building tiny AI models creates a challenge. They need to match larger models in quality while using a fraction of the computing power.

“It’s not enough to say, OK, I’ll make a small model,” Stock said. “What you need is a small model that has the same quality as larger models, right?”

Mistral claims their benchmarks show improved error rates compared to competitors. Of course, companies always tout their own benchmarks. Real-world testing will reveal whether these models truly deliver.

Both models are available now. Voxtral Realtime works through Mistral’s API and Hugging Face, where you can try a demo. Voxtral Mini Transcribe 2 is available via the API or in Mistral’s AI Studio.

Why On-Device Processing Matters

Cloud-based AI isn’t going anywhere. But on-device models address real privacy concerns. Your sensitive conversations don’t travel across the internet. They don’t sit in company databases. And they can’t be subpoenaed, hacked, or accidentally leaked.

Journalists interviewing confidential sources benefit. So do healthcare providers discussing patient information. Legal professionals can transcribe sensitive client meetings without worrying about cloud security breaches.

Speed is the bonus. No upload time. No processing delays. No waiting for results to download. Everything happens right where you need it.

The trade-off? These small models might not match the accuracy of massive cloud-based systems in every scenario. But for many users, that’s an acceptable compromise for privacy and speed.

Privacy-focused AI tools represent a growing trend. As people become more aware of data collection practices, demand for local processing will increase. Mistral’s new transcription models arrive at exactly the right time.

Just make sure you teach them your name correctly first.