Home Stories About Search RSS Feed
Models 2 min read

Microsoft Enters the Model Race with MAI-Transcribe, MAI-Voice, and MAI-Image

Back to News

Microsoft has launched three new MAI (Microsoft AI) models, available through Microsoft Foundry on Azure. The release represents Microsoft’s clearest signal yet that it intends to build — not just distribute — frontier AI capabilities.

The Three Models

MAI-Transcribe-1

A speech-to-text model supporting the top 25 most-used languages:

  • 3.9% average Word Error Rate on the FLEURS benchmark — state-of-the-art accuracy
  • 2.5x faster batch transcription than previous Azure offerings
  • Engineered for noisy, real-world audio environments

MAI-Voice-1

A high-fidelity speech generation model:

  • Produces 60 seconds of natural, expressive audio in under one second on a single GPU
  • Supports custom voice creation from just a few seconds of audio samples
  • Designed for enterprise voice applications, accessibility, and content creation

MAI-Image-2

A text-to-image model focused on quality and control:

  • Debuted at #3 on the Arena.ai leaderboard for image model families
  • Excels in natural lighting, accurate skin tones, and texture fidelity
  • Generates clear in-image text — a persistent weakness in competing models
  • An efficient variant, MAI-Image-2e, followed on April 14 for high-volume production workflows

Microsoft Foundry

All three models are available through Microsoft Foundry, the company’s central platform for model discovery, evaluation, fine-tuning, and deployment. The platform also hosts models from OpenAI, Meta, and other partners, with built-in responsible AI guardrails and governance controls.

Notably, Microsoft expanded Foundry Local into public preview the same week — allowing enterprise and sovereign customers to run AI workloads on-premises without cloud connectivity.

The Strategic Angle

These are the same models powering Copilot, Bing, and PowerPoint. By making them available as standalone APIs, Microsoft is:

  1. Reducing dependency on OpenAI for its own product stack
  2. Building a cost-competitive alternative for enterprise developers who don’t need frontier reasoning
  3. Creating a vertically integrated AI platform where Microsoft controls the full stack from model to application

The message is clear: Microsoft is no longer content to be just the infrastructure layer behind someone else’s models.


Source: microsoft.ai, microsoft.com, redmondmag.com

Marcus Chen
Written By

Marcus Chen

Lead Tech Analyst

Marcus is a hardware specialist and machine learning systems analyst who tracks large language model architectures, cloud compute infrastructure, and GPU accelerators. He specializes in decoding training efficiency and hardware benchmarks.