Microsoft unveils in-house MAIVoice1 and MAI1preview for Copilot

Microsoft on Thursday unveiled two AI models it said were developed entirely inhouse, a move that signals a strategic shift toward building consumer-focused capabilities independent of its longtime partner OpenAI. The new models, MAIVoice1 and MAI1preview, will be tested in public and begin appearing in select Copilot features in the coming weeks, Microsoft said.

MAIVoice1 is a speech generation model Microsoft says can produce a full minute of audio in under one second while running on a single GPU. The company said the model is already powering Copilot Daily, which delivers brief AI-narrated news summaries, and Copilot Podcasts, which generates podcast-style explainers from prompts. Microsoft demonstrated MAIVoice1 producing high-fidelity, expressive audio in both single- and multi-speaker scenarios and made the model available for hands-on trials in Copilot Labs.

MAI1preview is a text-focused foundation model Microsoft described as an instruction-following system optimized for consumer use cases. Microsoft said it trained MAI1preview on roughly 15,000 Nvidia H100 GPUs and is running inference on a single GPU, with experiments also underway on Nvidia GB200 hardware to boost...