Microsoft's AI division has unveiled its first in-house AI models, MAI-Voice-1 and MAI-1-preview. This move signifies a strategic shift towards internal AI development, potentially reducing reliance on external partners. MAI-Voice-1 excels in generating high-fidelity audio, producing a minute of natural-sounding speech in under a second using a single GPU. It supports various applications, including interactive assistants and podcast narration, with low latency.
MAI-1-preview, a foundational language model, was trained on Microsoft's infrastructure using 15,000 NVIDIA H100 GPUs. It is optimised for instruction-following and everyday conversational tasks. MAI-Voice-1 is integrated into Microsoft products like Copilot Daily for voice updates and news summaries. MAI-1-preview is undergoing public testing on LMArena. These models are designed for cost-effectiveness and scalability, positioning Microsoft as a strong contender in the AI market.
