The short answer
Choose MAI-Transcribe-2 if you already build on Azure and want 60 languages, speaker diarization, keyword biasing and a choice of clean or verbatim transcripts from one hosted call. Choose Phonon-2.5 if the audio must stay on your own hardware or offline, or you want a model with no per-hour fee that covers Croatian and Maltese, which are not in MAI-Transcribe-2's language table.
MAI-Transcribe has no row in the Open ASR Leaderboard snapshot we compare against and we have not scored it ourselves, so this page makes no accuracy comparison.
Side by side
Microsoft facts from Microsoft Learn (MAI-Transcribe in Azure Speech) and microsoft.ai/models/mai-transcribe-2, read 2026-11-13.
| Phonon-2.5 | MAI-Transcribe-2 | |
|---|---|---|
| Model type | Open weights on Hugging Face | Hosted API (Azure Speech, Microsoft Foundry) |
| Size | 156 MB download (Phonon-2.5 English: 158 MB) | Not published |
| Languages | 25 European languages (Phonon-2.5 English: English only) | 60, with automatic language identification; 23 of Phonon-2.5's 25 (not Croatian or Maltese) |
| English word error, seven Open ASR sets | 5.06 % (Phonon-2.5 English: 4.82 %) | No row in the snapshot; not measured by us |
| Price per audio hour | No per-hour fee when self-hosted | $0.10 (Microsoft lists it as a limited-time price to the end of the year); Streaming variant $0.54 |
| Where it runs | Your hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, Android | Azure regions listed by Microsoft |
| Streaming | Yes, live WebSocket from fermion serve | MAI-Transcribe-2 Streaming (no diarization or word timestamps) |
| Diarization | No | Yes (batch model) |
| Word timestamps | Yes (--json output) | Yes (batch model) |
| Transcript style | Verbatim | Verbatim or clean (fillers removed) |
Languages: where the two overlap
MAI-Transcribe-2's language table covers 23 of Phonon-2.5's 25 languages. Croatian and Maltese are in Phonon-2.5 and not in the table; Bosnian is in the table. MAI-Transcribe-2 adds Arabic, Chinese, Hindi, Japanese, Korean, Turkish and about thirty more, and handles code switching such as Hinglish and Spanglish.
Microsoft also lists MAI-Transcribe-1.5 (43 languages per OpenRouter) and notes that MAI-Transcribe-1 was deprecated on 20 August 2026.
Where MAI-Transcribe-2 is the better choice
- Azure-native: same subscription, keys and regions as the rest of Azure Speech.
- Speaker diarization and keyword biasing (phrase lists) in the batch model.
- A clean transcript style that removes fillers, as well as verbatim.
- 60 languages with automatic detection and mid-sentence code switching.
Where Phonon-2.5 is the better choice
- Runs outside any cloud: on-premises servers, laptops, Macs and Android devices, including with no network.
- No per-hour fee and no promotional price to expire.
- Croatian and Maltese.
- Open weights: the model you test is the model you ship, version for version.
pip install fermion-research
fermion transcribe phonon-2.5 interview.wav --json
fermion serve phonon-2.5 # POST /v1/audio/transcriptions, GET /v1/audio/streamQuestions
- Is Phonon-2.5 more accurate than MAI-Transcribe-2?
- We cannot say: MAI-Transcribe has no row in the Open ASR Leaderboard snapshot we use and we have not measured it with the same protocol. Run both on a sample of your own audio.
- How much does MAI-Transcribe-2 cost?
- Microsoft's model page lists MAI-Transcribe-2 at $0.10 per audio hour as a limited-time price to the end of the year, and the Streaming variant at $0.54 per hour (read 2026-11-13).
- Which languages does MAI-Transcribe-2 support?
- Sixty, per Microsoft Learn. Of Phonon-2.5's 25 languages, all except Croatian and Maltese are in its table.
- Can MAI-Transcribe run on my own hardware?
- Microsoft documents it as a model called through Azure Speech in Azure regions. Phonon-2.5 is open weights and runs on your own machines.
- Does Phonon-2.5 remove filler words like MAI-Transcribe's clean style?
- No. Phonon-2.5 transcribes verbatim, keeping repetitions and fillers. MAI-Transcribe-2 offers both verbatim and clean styles.
Related
Sources and method
- Languages, features, model versions and deprecation: learn.microsoft.com/azure/ai-services/speech-service/mai-transcribe, read 2026-11-13.
- Prices, Streaming variant features: microsoft.ai/models/mai-transcribe-2, read 2026-11-13. MAI-Transcribe-1.5 language count: OpenRouter model pages, read 2026-11-11.
- Phonon-2.5 and Phonon-2.5 English: Fermion Research measurement, Phonon-2.5 release post (2026-10-22), Open ASR Leaderboard scoring code, full test sets of the same seven sets.
Canonical: https://www.fermionresearch.com/models/phonon-2-5/vs-microsoft-mai-transcribe/