Microsoft’s AI division introduced its first homegrown AI fashions on Thursday: MAI-Voice-1 AI and MAI-1-preview. The corporate says its new MAI-Voice-1 speech mannequin can generate a minute’s price of audio in beneath one second on only one GPU, whereas MAI-1-preview “provides a glimpse of future choices inside Copilot.”
You’ll be able to attempt MA1-Voice-1 out for your self on Copilot Labs, the place you’ll be able to enter what you need the AI mannequin to say, in addition to change its voice and elegance of talking. Along with this mannequin, Microsoft launched MAI-1-preview, which it says it skilled on round 15,000 Nvidia H100 GPUs. It’s constructed for customers in want of an AI mannequin able to following directions and “offering useful responses to on a regular basis queries.”
Microsoft AI chief Mustafa Suleyman mentioned throughout an episode of Decoder final yr that the corporate’s inner AI fashions aren’t targeted on enterprise use instances. “My logic is that we’ve to create one thing that works extraordinarily properly for the patron and actually optimize for our use case,” Suleyman mentioned. “So, we’ve huge quantities of very predictive and really helpful knowledge on the advert aspect, on client telemetry, and so forth. My focus is on constructing fashions that actually work for the patron companion.”
Microsoft AI plans on rolling out MAI-1-preview for sure textual content use instances in its Copilot AI assistant, which presently depends on OpenAI’s massive language fashions. It has additionally began publicly testing its MAI-1-preview mannequin on the AI benchmarking platform LMArena.
“We’ve got large ambitions for the place we go subsequent,” Microsoft AI writes within the weblog publish. “Not solely will we pursue additional advances right here, however we consider that orchestrating a spread of specialised fashions serving completely different person intents and use instances will unlock immense worth.”
