Microsoft has expanded its MAI speech model line with three new solutions: MAI-Transcribe-2-Streaming, MAI–Voice-2.1, and MAI–Voice-2.1-Flash. The updates focus on real-time speech recognition, low-latency voice generation, and building conversational AI assistants.
Microsoft has expanded its MAI speech model line with three new solutions: MAI-Transcribe-2-Streaming, MAI–Voice-2.1, and MAI–Voice-2.1-Flash. The updates focus on real-time speech recognition, low-latency voice generation, and building conversational AI assistants.
The main news of the release was the MAI-Transcribe-2-Streaming model. Unlike the basic MAI-Transcribe-2, which works with recorded audio, the streaming model updates the transcript directly while a person is speaking, WindowsReport reports .
Initial transcription results appear just over 100 ms after the audio signal is received, allowing for the creation of voice agents that respond almost instantly.
The model supports 60 languages and independently determines the speaker's language during a conversation without the need for prior configuration by developers.
According to the Artificial Analysis platform, the model ranked first in terms of accuracy for both partial and final transcription.
The discounted usage fee is $0.54 per hour of audio until the end of 2026.
Microsoft also introduced an updated speech generation model, MAI–Voice-2.1, designed for creating multilingual voice applications.
The model covers 23 languages (for 26 regional settings) and is able to switch between them while maintaining the unique identity and timbre of a single speaker.
To reproduce the voice of a specific person in any of the supported languages of the model, only a few seconds of audio recording are enough.
The generation price is $22 per 1 million characters.
For tasks where speed and scalability play a key role, the MAI–Voice-2.1-Flash model has been released.
The model is capable of generating 45 seconds of audio with a latency of about 150 ms. It provides 55% faster inference and costs about 60% less than its counterparts — $15 per 1 million characters. At the same time, it retains voice cloning capabilities.
The combination of streaming recognition and speech synthesis gives developers a complete technology stack from Microsoft for building real-time conversational AI agents.
All three new models are available now through Microsoft Foundry, MAI Playground, and Vercel. The MAI–Voice-2.1 and MAI–Voice-2.1-Flash models are also integrated into OpenRouter, with LiveKit support coming soon. They can also be used through the Azure Voice Live platform.
