Microsoft unveils 3 new AI models for speech, voice, image

Microsoft has launched three new MAI AI models – MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 – and made them available to developers through Microsoft Foundry, with the MAI Playground (the US only) offering a place to try them.

Microsoft said MAI-Transcribe-1 targets speech to text across the top 25 most-used languages and scored a 3.9% average word error rate on the FLEURS benchmark, and it said batch transcription runs 2.5 times faster than its Azure Fast offering.

Microsoft also said MAI-Voice-1 can generate a custom voice from a few seconds of audio and produce 60 seconds of audio in about one second.

Microsoft said MAI-Image-2 is on Copilot, with phased rollouts underway in Bing and PowerPoint, with pricing starting at US$0.36 per hour for transcription, US$22 per 1 million characters for voice, and US$5 per 1 million text tokens plus US$33 per 1 million image tokens for image output.

🔗 Source: Microsoft

🧠 Food for thought

Implications, context, and why it matters.

Microsoft’s new models mark a careful shift away from OpenAI reliance

The release came after a contract renegotiation that, until October 2025, had barred Microsoft from independently pursuing artificial general intelligence (AGI), or human-level AI 1.
Microsoft now says it wants “AI self-sufficiency,” building in-house models to gain leverage and cut AI-related costs 1.
Small Microsoft AI engineering teams with fewer than 10 people built the models, with leaders saying the MAI-Transcribe-1 transcription model uses nearly half the GPU (graphics processing unit) resources of state-of-the-art rivals 1.
On hardware, Microsoft has built custom AI chips called Maia to run its AI workloads 2.

Special-purpose AI features are starting to look like commodities

Microsoft offers the models through Microsoft Foundry at aggressive prices, treating specialized AI features like low-cost utilities 1.
This move squeezes venture-backed startups that sell single-function tools such as voice synthesis or transcription, since they now face a cheaper, bundled option from a tech giant 1.
Microsoft courts cautious corporate clients by stressing a “clean lineage” of models plus data provenance (clear records of where training data came from), which can matter amid copyright lawsuits facing many AI labs 1.
By listing its models next to options from partners like OpenAI and Anthropic, Microsoft frames Foundry as what it calls a “platform of platforms” for developers 1.