Microsoft unveils its first homegrown AI models MAI-Voice-1 and MAI-1-Preview
Microsoft is already deploying the model in products such as Copilot Daily, which delivers narrated news briefings, and in podcast-style explainers designed to make complex topics more accessible.
ADVERTISEMENT
Microsoft's AI division on Thursday announced its first in-house AI models - MAI-Voice-1 AI and MAI-1-preview - signalling a significant milestone in the tech giant's pursuit of building proprietary artificial intelligence to power its Copilot platform.
The MAI-Voice-1 speech model stands out for its speed and efficiency, capable of generating one minute of audio in less than a second using just a single GPU.
Microsoft is already deploying the model in products such as Copilot Daily, which delivers narrated news briefings, and in podcast-style explainers designed to make complex topics more accessible.
Users can experiment with MAI-Voice-1 on Copilot Labs, where they can generate customized speech with different voices and speaking styles.
Alongside this, Microsoft also introduced MAI-1-preview, a model trained on roughly 15,000 Nvidia H100 GPUs.
Built to handle instruction-following and everyday queries, it offers an early glimpse into the company's future AI roadmap. The model is already undergoing testing on the benchmarking platform LMArena, with plans to integrate it into select Copilot text-based features.
Also Read: Microsoft AI Chief warns of public push for ‘AI rights’ as users grow attached to chatbots
Mustafa Suleyman, Microsoft AI Chief, reiterated that the company's approach is consumer-first rather than enterprise focused.
Speaking previously on Decoder, he explained, "We have to create something that works extremely well for the consumer and really optimize for our use case."
He emphasized Microsoft's unique advantage in leveraging data from consumer behaviour and advertising to train models that can act as true digital companions.
In a blog post, Microsoft wrote: "We have big ambitions for where we go next. Not only will we pursue further advances here, but we believe that orchestrating a range of specialized models serving different user intents and use cases will unlock immense value."
