Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Sometimes you want to copy something, but don’t want it hanging online for a hacker to see. Maybe it’s a conversation with your doctor or lawyer. Maybe you’re a journalist, and it’s a sensitive interview. Privacy and control are important.
This desire for privacy is one of the reasons French developer Mistral AI built its latest clone model to be small enough Run on devices. It can be played on your phone, on your laptop, or in the cloud.
Foxtral Mini Transcript 2One of the new models announced Wednesday is “very, very small,” Pierre Stock, Mistral’s vice president for science operations, told me. Another new model, Voxtral Realtime, can do the same thing but is as straightforward as closed captioning.
Privacy isn’t the only reason the company wants to build small businesses Open source models. By running directly on the device you’re using, these models can run faster. No need to wait for files to find their way over the Internet to the data center and back.
“What you want is for the transcription to be done nicely, very close to you,” Stock said. “The closest thing we can find to you is any high-end device, like a laptop, a phone, or a wearable device like a smartwatch, for example.”
Low latency (read: high speed) is especially important for real-time transcription. The Voxtral Realtime model can generate with latency of less than 200 milliseconds, Stock said. It can transcribe a speaker’s words as quickly as they can be read. No more waiting two or three seconds for closed captions to continue.
Watch this: Chip shortage hits iPhones, OpenAI investment halts, TikTok censorship allegations | Technology today
The Voxtral Realtime model is available through the Mistral API and On the face hugalong with Demo Where you can try it.
In some brief testing, I found it generated fairly quickly (although not as quickly as you’d expect if it were on the device) and it was able to accurately pick up what I said in English with a little Spanish mixed in. It is capable of handling 13 languages at the moment, according to Mistral.
Voxtral Mini Transcribe 2 is also available through the company’s API, or you can play with it Mistral Artificial Intelligence Studio. I used the template to transcribe my interview with Stock.
I found it to be very fast and reliable, although it had difficulty using correct names like Mistral AI (which I call Mr. Lay Eye) and Voxtral (VoxTroll). Yes, the AI model got it wrong. But Stock said users can customize the model to better understand certain words, names and terms if they are using it for specific tasks.
The challenge of building small, fast AI models is that they must also be accurate, Stock said. The company touted the models’ performance against benchmarks that showed improved error rates compared to competitors.
“It’s not enough to say, ‘Okay, I’ll make a little model,'” Stock said. “What you need is a small model that is just as good as the larger ones, right?”