Microsoft just put out three new AI models built for one job: understanding and speaking human language in real time, not after a pause. The headline one is called MAI-Transcribe-2-Streaming, and it is Microsoft’s first model built to turn speech into text while you are still talking, not after you stop.

The numbers behind it are honestly wild. In independent testing by Artificial Analysis, it came out on top among 38 models for accuracy, with a word error rate of about 2.5 percent. It gets you a first readable chunk of text in a little over a tenth of a second, and it handles 60 languages.

Alongside it, Microsoft released two text-to-speech models, MAI-Voice-2.1 and a faster version called MAI-Voice-2.1-Flash. The bigger one can speak the same voice across 23 languages, each with a native-sounding accent instead of a flat, translated-over tone.

Speed is the real story here. Anything built to actually talk with you, a customer service line, a voice assistant, a live captioning tool, falls apart the moment there is a lag. People notice a half-second delay in conversation. These models are built to close that gap so the back and forth feels normal instead of robotic.

Pricing is introductory for now, under a dollar for an hour of transcription, which tells you Microsoft wants this used, not just admired in a demo. Voice tech has been the quiet corner of the AI race while image and text models get the spotlight. This release is a reminder that the quiet corner just got a lot more competitive, and that is good news for anyone who has ever been stuck waiting on a call center bot to catch up with what they said.

——

Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277

——

Sources:
https://microsoft.ai/news/our-first-streaming-transcription-model/
https://the-decoder.com/microsoft-ai-releases-new-transcription-and-text-to-speech-models-for-voice-agents/
https://datanorth.ai/news/microsoft-releases-mai-transcribe-2-streaming

Repost this. Thanks.