Microsoft launches its own voice models: AI that listens and speaks at once

Microsoft is expanding its MAI family of in-house AI models with more voice models. They are meant to make it easier to build voice-controlled assistants and customer service that can speak multiple languages.

MAI Realtime: listens and speaks at once

The most eye-catching model is MAI Realtime. According to coverage of the model, it is Microsoft's first native real-time voice model developed in-house. It is so-called "full-duplex". That means it can listen and speak at the same time instead of taking turns, as when two people use walkie-talkies. In practice, an AI can thus let itself be interrupted and respond more naturally in a conversation.

The model supports 17 languages and has two starter voices, Victoria and Grant. Access is so far limited to selected partners, so ordinary users cannot just try it yet.

Fast speech and transcription

MAI-Voice-2-Flash is a text-to-speech model designed to deliver fast and natural speech in 15 languages. It is optimized for low latency, so answers can be heard almost immediately. That matters in voice assistants, where even short pauses feel awkward.

MAI-Transcribe-1 goes the other way and turns speech into text in up to 25 languages. The list includes Danish, Swedish, German and French, among others. According to coverage, Microsoft claims the model uses about half as much GPU power as leading competitors. A GPU is the type of chip that powers AI computations. That figure is Microsoft's own and has not been independently verified in what I have been able to find.

According to the sources, the older MAI-Voice-1 can generate a minute of audio in under a second on a single GPU.

What does it mean for you?

For ordinary users, the effect will come indirectly: voice assistants in Copilot and in companies' apps may become faster and more natural to talk to. That Danish is included in the transcription model is good news for things like automatic meeting minutes and subtitles.

For Microsoft, it is also about becoming less dependent on external suppliers such as OpenAI by building its own models. Whether MAI Realtime will support Danish does not appear in what I have been able to find.

What it means for you

If you use voice assistants or call customer service, you may eventually experience AI that answers faster and lets itself be interrupted, as a real person would. Danish is included in the model that turns speech into text, so keep an eye out for tools for automatic meeting minutes and subtitles built on it. You cannot try MAI Realtime yourself yet, since only selected partners have access, so right now there is nothing you need to do other than see what your current tools update to.