Microsoft's new streaming transcription starts guessing in about 100ms. Voice agents can now think while callers talk

The awkward pause is what gives most AI phone agents away. Microsoft's new models go straight at it.

What happened

Microsoft AI released MAI-Transcribe-2-Streaming, a real-time transcription model. Microsoft says it ranks first for accuracy on Artificial Analysis. It covers 60 languages and, according to The Decoder, delivers its first partial results in just over 100 milliseconds, so a voice agent can start responding while someone is still mid-sentence. The introductory price is $0.54 per hour of audio through the end of the year.

Two text-to-speech models came with it. MAI-Voice-2.1 can speak 23 languages in the same voice with a native accent in each. MAI-Voice-2.1-Flash targets speed, with Microsoft citing 150 milliseconds of latency at $15 per million characters instead of $22 for the standard model. Both can clone a voice from a few seconds of reference audio, with built-in safeguards meant to prevent misuse. They are available through Microsoft Foundry and the MAI Playground, and the two voice models are also on OpenRouter. In one test, about half of 4,000 participants thought the voices belonged to a real person.

My take

For anyone building phone or support automations, three things stand out.

First, latency is the real UX metric. Partial transcripts in about 100ms mean your agent can start looking up the caller's record in the CRM before they finish the sentence. That is where a voice agent stops feeling like a phone tree.

Second, one voice across 23 languages matters for businesses with global customers. Many of my clients serve callers in more than one language, and switching voices mid-call confuses people.

Third, the "half thought it was human" result is a reason for more disclosure, not less. If callers cannot tell, say it plainly at the start of the call and log the conversation.

The model is now rarely the bottleneck. The work is connecting the voice layer to your booking system, CRM and escalation rules so the agent can actually resolve the call.

If you have a phone task your team dreads, like missed call follow-ups or appointment confirmations, message me about it through romielwillautomate.dev and I will tell you honestly whether a voice agent fits.