MAI-Transcribe-2 Online
Transcribe audio with Microsoft's MAI-Transcribe-2 in your browser — no Azure account, no API key, no code. Speaker labels and timestamps included.
What is MAI-Transcribe-2?
MAI-Transcribe-2 is the second generation of Microsoft AI's own speech-to-text model, released on September 3, 2026. According to its model card, it covers 60 languages with automatic language detection, and this generation adds speaker diarization, word-level timestamps and keyword biasing for names and jargon.
Microsoft reports the lowest average word error rate across the 60 languages of the FLEURS benchmark among the systems it compared, and second place on the Artificial Analysis English speech-to-text leaderboard. It also claims about 10 seconds to transcribe an hour of audio.
Developers reach it through the Azure Speech fast transcription API, where it is a public preview. ScribeForge runs those calls for you: you drop a file and read the transcript.
How to use it — 3 steps
- Open ScribeForge with MAI-Transcribe-2 selected. The engine menu above the button says "MAI-Transcribe-2 (Microsoft)".
- Drop your audio or video file. MP3, WAV, FLAC, M4A, MP4, OGG, Opus, WebM, AAC, MOV and more: formats Azure does not take are converted to MP3 first.
- Read, copy or export the transcript with speaker labels and timestamps. Click a segment to jump the audio there.
We ran it next to Grok STT
ScribeForge has two engines, so on October 9, 2026 we transcribed the same file with both, three times each: the 6 min 17 s LibriVox recording of Mark Twain's A Telephonic Conversation (the demo file on the home page).
| MAI-Transcribe-2 | Grok STT (xAI) | |
|---|---|---|
| Time for 6 min 17 s (median of 3) | 4.7 s | 3.7 s |
| Words returned | 826 | 815 |
| Speakers labelled | 2 | 1 |
| Numbers | digits ("50") | words ("fifty") |
| The word "Valenciennes" | "Valenceins" | "Valenciennes" |
The two transcripts agree on 95.9% of the words; most differences are Twain's nonsense sounds ("clook" vs "cluck", "lolly" vs "roly"). Times include the upload from our server and are not a benchmark: one file, one day. If one engine is down or refuses a file, ScribeForge sends it to the other automatically.
MAI-Transcribe-2 vs other ways to run it
| Option | What you need | Good for |
|---|---|---|
| ScribeForge | A browser | Transcribing files now, with exports and speaker labels |
| Azure Speech API | Azure subscription, resource, key, code | Building it into your own app |
| MAI Playground | A browser, on Microsoft's demo site | Trying the model on Microsoft's own page |
| OpenRouter | OpenRouter account and credits, code | Calling many models through one API |
Common questions
MAI-Transcribe-2 is Microsoft AI's in-house speech-to-text model, released on September 3, 2026. It transcribes 60 languages with automatic language detection, and adds speaker diarization and word-level timestamps. It is offered through Azure Speech as a public preview.
Not with ScribeForge. ScribeForge calls MAI-Transcribe-2 through its own Azure resource, so you upload a file in the browser and get the transcript without an Azure subscription, API key or code.
Azure accepts WAV, MP3 and FLAC. ScribeForge converts other formats (M4A, MP4, OGG, Opus, WebM, AAC, MOV and more) to MP3 before sending them, up to 100 MB per file.
No. The file is held in server memory, sent to Azure Speech in Microsoft's North Europe region for transcription, and discarded when the request ends.
You get 2 free previews a day: the start of the transcript, from the first minute of audio. For full transcripts, a pack of 50 is $9 once (never expires) or unlimited is $19/month. See MAI-Transcribe-2 pricing for what the API itself costs.
Transcribe with MAI-Transcribe-2 — no Azure account needed.
Open MAI-Transcribe-2 →No account · No API key · No credit card · 2 free previews a day
Related
ScribeForge is an independent tool and is not affiliated with Microsoft. MAI-Transcribe is a Microsoft model; facts on this page come from Microsoft's model card and Azure documentation, checked October 9, 2026.