Microsoft AI Unveils New Transcription and Text-to-Speech Models for Voice Agents

Meta description: Microsoft AI releases MAI-Transcribe-2-Streaming for real-time transcription, plus MAI-Voice-2.1 models, with just over 100 ms latency and 60 languages.

Tags: Microsoft AI, real-time transcription, MAI-Transcribe-2-Streaming, voice agents, Artificial Analysis

Featured image: Abstract technology style image of streaming audio waveforms and multilingual text flowing into a voice agent pipeline, with no real people, no company logos, and no text.

Microsoft AI real-time transcription moved closer to live conversation with the launch of MAI-Transcribe-2-Streaming, a model designed to turn speech into text while a speaker is still talking. The company says the model handles 60 languages and delivers its first partial results in just over 100 milliseconds, a timing detail that matters for voice agents that need to acknowledge or answer before a user finishes a sentence. Alongside it, Microsoft released two new text-to-speech models, including MAI-Voice-2.1, giving developers a more complete pipeline for listening, understanding, and speaking.

Background: The Voice Agent Bottleneck

For an AI system that takes spoken input, the practical challenge is not only recognizing words but doing so quickly enough for the exchange to feel like a conversation. Traditional batch transcription can wait until an utterance ends, then return a final transcript. In live agent workflows, that pause can break turn-taking: the user stops speaking, waits for the system to process, and the system then replies. Microsoft says MAI-Transcribe-2-Streaming lets voice agents respond while someone is still mid-sentence, because first partial results arriving in just over 100 milliseconds can begin the reply path before the utterance is complete.

The text-to-speech side has a similar requirement. A generated response may be accurate but unnatural if it sounds like one fixed voice reading a generic accent for every language. Microsoft says MAI-Voice-2.1 is designed to speak 23 languages in the same voice, with a native accent in each one. That positions the models as a pair: fast transcription for input, and consistent synthetic speech for output.

Development: From Streaming Transcription to Voice Synthesis

Microsoft says MAI-Transcribe-2-Streaming ranks first for accuracy on Artificial Analysis, a benchmark platform that compares speech recognition models. The ranking claim is part of the release narrative, but the more operational detail is the multilingual scope and latency target. With 60 languages covered, the model can be used in products that need one transcription layer across many markets, rather than separate regional speech models.

The release also expands Microsoft’s MAI voice family. MAI-Voice-2.1 is intended for speech synthesis across 23 languages while preserving a single speaker identity. The company further highlights MAI-Voice-2.1-Flash as a lower-latency, lower-cost variant. According to Microsoft, the Flash variant hits a latency of 150 milliseconds and costs $15 per million characters instead of $22. That difference suggests a trade-off aimed at high-volume or interactive applications where both speed and unit economics matter.

Both voice models can clone a voice from just a few seconds of reference audio. Microsoft says built-in safeguards are meant to prevent misuse, a necessary detail for any voice cloning tool because cloned audio can be convincing. In one test, about half of the 4,000 participants thought the voices belonged to a real person.

Latest Progress: Pricing, Benchmarks, and Availability

The models are available through Microsoft Foundry and the MAI Playground, among other platforms, and the two voice models are also on OpenRouter. For MAI-Transcribe-2-Streaming, Microsoft gives a clear introductory pricing window: through the end of the year, an hour of audio costs $0.54 at the introductory price. That price point helps developers estimate costs for continuous transcription workloads, such as meeting agents, support calls, or live captions.

Artificial Analysis ranking, 60-language coverage, and partial results in just over 100 ms form the headline package, while the voice side adds 23-language synthesis, a 150-millisecond Flash claim, and the $15 versus $22 per million characters pricing comparison. Together, these numbers outline a current stage where Microsoft AI is trying to compete on the full real-time voice loop rather than isolated transcription or synthesis capabilities.