Microsoft released three new foundational models on April 2 — MAI-Transcribe-1 for speech-to-text, MAI-Voice-1 for voice generation, and MAI-Image-2 for image creation. All three are available immediately through Microsoft Foundry and a new MAI Playground. They are built entirely in-house, without OpenAI’s involvement.
That last detail is the one that matters most. Microsoft has invested tens of billions in OpenAI, but these models represent capabilities the company developed on its own — a deliberate expansion of its proprietary AI stack into territory where it previously relied on partners or third-party solutions.
The Models
MAI-Transcribe-1 is Microsoft’s first-generation speech recognition model. It supports 25 languages, runs at 2.5x the batch transcription speed of Microsoft’s previous Azure Fast offering, and logs a word error rate of 3.9% — ahead of both Gemini 3.1 Flash and OpenAI’s GPT-Transcribe on the same benchmarks. Microsoft claims approximately 50% lower GPU cost compared to leading alternatives. Pricing starts at $0.36 per hour.
MAI-Voice-1 generates high-fidelity speech. The headline number: 60 seconds of expressive audio in under one second on a single GPU. It preserves speaker identity across long-form content and supports custom voice creation from a few seconds of reference audio. Pricing is $22 per million characters.
MAI-Image-2 is Microsoft’s upgraded text-to-image model. It debuted at #3 on Arena.ai’s image model leaderboard and runs at least twice as fast as its predecessor. Pricing is $5 per million input tokens.
| Model | Capability | Pricing | Key Metric |
|---|---|---|---|
| MAI-Transcribe-1 | Speech-to-text (25 languages) | $0.36/hour | 3.9% WER, 2.5x faster batch |
| MAI-Voice-1 | Voice generation | $22/M characters | 60s audio in <1s |
| MAI-Image-2 | Text-to-image | $5/M input tokens | #3 Arena.ai leaderboard |
The Competitive Positioning
The pricing is aggressive, particularly on transcription. At $0.36 per hour, MAI-Transcribe-1 undercuts most enterprise speech-to-text providers while delivering accuracy that competes with the best available models. For organizations running transcription at scale — call centers, legal discovery, media production — the cost differential is significant.
MAI-Voice-1 enters a market where ElevenLabs has set the quality bar and OpenAI’s voice capabilities are embedded in ChatGPT. Microsoft’s angle is enterprise integration: custom voice creation through Foundry, speaker identity preservation, and the compliance and data residency guarantees that come with the Azure ecosystem. The quality will need to match ElevenLabs to win creative use cases, but for enterprise workflows where integration with existing Microsoft infrastructure matters more than voice quality at the margins, the value proposition is clear.
MAI-Image-2’s #3 Arena.ai ranking puts it in competitive range with the current leaders. The speed improvement over its predecessor matters for production workflows where image generation is a pipeline step rather than a creative exploration.
What This Signals
The strategic significance is less about the individual models and more about the pattern. Microsoft is building multimodal AI capabilities that do not depend on OpenAI. This is not a hedge — it is a parallel track.
The relationship between Microsoft and OpenAI remains commercially important to both companies. Microsoft distributes OpenAI models through Azure, and the GPT-5.4 lineup remains the backbone of many Azure AI services. But with OpenAI’s valuation now at $852 billion and its ambitions expanding into media, Microsoft clearly sees value in owning more of its own stack. But Microsoft’s own model development — which also includes the Phi family for smaller, efficient models — is clearly expanding in scope and ambition.
For enterprise customers evaluating their AI infrastructure decisions, the practical implication is more options within the Microsoft ecosystem. A company already running on Azure can now build voice-enabled applications, transcription pipelines, and image generation features using Microsoft’s own models, reducing dependency on any single model provider. Whether to use MAI-Transcribe-1 or OpenAI’s Whisper, MAI-Image-2 or DALL-E, becomes a performance and pricing decision rather than a platform commitment.
The Multimodal Shift
These releases also reflect a broader industry trend: the focus of AI competition is expanding beyond text. The large language model race has dominated headlines for three years, but the real-world applications driving enterprise adoption increasingly require speech, voice, and vision capabilities. Microsoft’s bet is that owning these building blocks — not licensing them — gives it a stronger position as multimodal AI moves from experimental to essential.
The MAI Playground, where developers can test all three models interactively, is a smart distribution play. Lowering the friction between discovery and integration is how platforms win developers, and developers are how platforms win enterprises. Microsoft has played this game before.