Microsoft Unveils AI That Transcribes While You Speak, Plus Korean-Capable Voices

Microsoft on October 1 (local time) unveiled MAI-Transcribe-2-Streaming, an AI model that turns speech into text while a person is still talking. The company also released two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. Developers can access all three through Microsoft Foundry and the MAI Playground.

Transcribing mid-sentence: how fast is it?

The new transcription model displays text as someone is speaking. It automatically recognizes 60 languages, and as more speech arrives it revises what it has already written before finalizing the result. That differs from the conventional approach of transcribing an entire recording after it ends.

Microsoft says the model begins producing interim results roughly 0.1 seconds after audio comes in. In comparisons the company published, words typically appeared on screen about 0.32 seconds after being spoken. The closest competing model took more than 0.5 seconds, according to Microsoft. The company also says the model ranked first on Artificial Analysis accuracy evaluations for both interim and final results.

The existing MAI-Transcribe-2 converts already-recorded audio into text, identifies who spoke, and can time-stamp individual words. What sets the new Streaming model apart is that results arrive while speech is still in progress. Language coverage has also grown from 43 in MAI-Transcribe-1.5 to 60.

Two text-to-speech models: does Korean work?

The MAI-Voice-2.1 speech synthesis model supports 23 languages and 26 locales. Its predecessor, MAI-Voice-2, covered 15 languages. The new model keeps the same speaker’s voice intact even when the language changes, which lets services that generate guidance audio in multiple languages maintain a single consistent voice.

Flash supports the same languages as the standard model but is tuned for fast responses and high-volume workloads. Microsoft says Flash delivers 55% faster model inference than the comparison model, at roughly 60% lower cost. Korean appears on the official voice lists for both models, so developers looking to add Korean-language speech can test it.

What it costs and where it runs

Introductory pricing for the transcription model is $0.54 per hour of audio through the end of 2026. Speech synthesis costs $22 per million characters for the standard model and $15 for Flash. Microsoft says the speech synthesis models are available worldwide. However, Azure’s Korea regions are absent from the list of regions that process requests.

All three models are in public preview. Microsoft’s documentation notes that no service-level agreement applies and that production use is not recommended. Voice cloning uses a 5- to 60-second reference recording for which the speaker has given consent, and it is available only after a separate review process.

With this release, developers can test live transcription and multilingual speech output side by side. Anyone looking to build these into services in Korea will need to weigh both the transcription model’s Korean coverage and where data is processed. Accuracy and response speed in real Korean conversations, along with the terms for moving from preview to general availability, remain to be confirmed.