Microsoft released three in-house AI models focused on transcription and voice. The models are MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash. All three became available right away. Microsoft AI described the release as accurate streaming transcription paired with natural speech and less waiting between turns. The company is building more of its own AI technology as it works to attract developers who create conversational voice agents.

MAI-Transcribe-2-Streaming Features and Performance

MAI-Transcribe-2-Streaming is Microsoft’s first streaming transcription model. It converts speech into text while a person is still talking. It doesn’t wait for the full recording to finish. The model supports 60 languages. It can also detect which language is being spoken automatically, even when the language changes during a conversation. Its introductory price is $0.54 per hour of audio.

The model ranked first on Artificial Analysis’s AA-WER Streaming benchmark, which measures word error rate. It led in both final transcript accuracy and first partial transcript accuracy, posting a 2.5% word error rate. Microsoft also states that the model debuted at number one on Artificial Analysis.

This release builds on MAI-Transcribe-2. Microsoft’s documentation describes MAI-Transcribe-2 as a next-generation speech-to-text model built in-house by the Microsoft AI team. Developers could already access that earlier model through Azure’s speech service.

MAI-Voice-2.1 and MAI-Voice-2.1-Flash

MAI-Voice-2.1 produces expressive, low-latency speech that remains stable across long passages. OpenRouter’s model listing calls it Microsoft AI’s highest-fidelity text-to-speech model. It works in 23 languages. MAI-Voice-2.1-Flash is a lighter version designed for speed. Both voice models, along with the streaming transcriber, are available through Vercel’s AI Gateway at Microsoft’s listed rates.

Microsoft’s Expanding MAI Model Family

The new models form the latest addition to the MAI family from Microsoft AI. That group previously released the MAI-Voice and MAI-1 models. In a market note, 24/7 Wall St. pointed out that each layer of the AI stack Microsoft builds itself is one where it no longer depends on an outside supplier. The outlet said the benchmark result suggests Microsoft’s internal capabilities are maturing faster than the public narrative has reflected.