Live transcription for Korean meetings: MAI-Transcribe-2-Streaming
MAI-Transcribe-2-Streaming claims 60 languages with auto detection and about 100ms to first text. Verify Korean yourself. Sume transcribes files, not live.

For a live Korean meeting, MAI-Transcribe-2-Streaming is worth a trial, but I could not confirm Korean support from the page I read. Microsoft's Oct 1, 2026 post says 60 languages with automatic language detection and first hypotheses in just over 100ms. It does not, in the text I read, list the 60 languages, so check that Korean is on the list before you plan around it. For meeting recordings after the fact, Sume can transcribe a hosted clip with a Korean language hint.
What Microsoft states
| Item | Stated |
|---|---|
| Languages | 60, with automatic language detection |
| Speed | First hypotheses in just over 100ms |
| Ranking claim | No. 1 on Artificial Analysis for final and partial transcripts |
| Price | $0.54 per hour, introductory through end of year |
| Availability | OpenRouter, Microsoft Foundry, MAI Playground, Vercel; LiveKit coming soon |
A Korean meeting test
Mixed-language meetings are where auto detection is tested. A Korean meeting often drops English product names, acronyms and numbers into the sentence. Before you pick a tool, record a real 10-minute excerpt and score it:
- Count wrong or missing proper nouns and product names per 100 words.
- Check how numbers, dates and units are written.
- Look at partial transcripts as well as final ones: a live caption that rewrites itself often is hard to read.
- Test two speakers who talk over each other, and a quiet speaker.
- Try the same clip with the language set explicitly and with auto detection.
The Sume route for recordings
Sume does not do live streaming transcription. POST /v1/video-inspect with transcribe: true reads one clip already on media.sume.com (import it first with POST /v1/media-imports) up to 1,800 seconds. Send language_code as a hint (the docs give ko as an example) or omit it to auto-detect, and segmentation.mode: "sentence" for gap-free sentence segments.
Cost for a 30-minute recording: 30 x $0.01 = $0.30 for the transcript, plus the job's compute, and a 5.5% platform fee on customer spend. MAI's introductory live rate for the same 30 minutes is 0.5 x $0.54 = $0.27. They are different products: one is live, the other works on a finished file.
Sources
Related posts
More in Use cases
- Make a portrait painting talk: H3 Max lip sync, 12 s for $1.20
A museum label, a family portrait or a book cover can speak: H3 Max lip sync takes a still and 5 to 14.8 seconds of audio. At 768p a 12 second clip is $1.20.
- Make an AI voice read an email or order code right: spell, verify
Amazon and Cartesia both claim better codes, emails and phone numbers. The reliable fix is text prep plus a read-back check. Try both on Sume's TTS router.
- Marketplace listing video from one product photo: first_frame
Turn a single product photo into a short listing video by pinning it as the first frame on seedance-2, then check the catalog fields that gate the request.
- Mascot dance ad: Kling motion control, orientation and sound
Drive a still mascot with a motion reference video through Kling 3.0 Motion Control on Sume. Settings that matter: duration, orientation and original sound.
Written by Sume