Microsoft debuts MAI-Transcribe-2-Streaming at 54 cents per audio hour

Microsoft's new models mark a step in its effort to build voice agents that can converse naturally with people. The company already released MAI-Transcribe-2 last month, and the streaming variant adds continuous, real-time output rather than waiting for a speaker to finish. Alongside the transcription model, Microsoft introduced MAI-Voice-2.1 and MAI-Voice-2.1-Flash for generating speech, priced at $22 and $15 per million characters respectively.

The company says a complete voice agent needs three capabilities: understanding speech, deciding what to do, and generating an audible reply. MAI-Transcribe-2-Streaming handles the first, the MAI-Voice models handle the third, and Microsoft's Mai-Thinking-1 reasoning model sits in the middle. Using separate models, according to Microsoft, gives developers more control over quality, latency and cost.

The releases also signal Microsoft's move away from model providers it invests in. Suleyman's comments to Bloomberg about reducing costs to Anthropic suggest the MAI family is intended to eventually power Copilot agents in platforms like Excel and Outlook, though how quickly that transition proceeds may depend on whether the in-house models can match the performance of the outside providers.