← All models

MAI-Transcribe

Audio Preview Full comparison ↗

microsoft-azure/mai-transcribe · by Microsoft AI · llm-based-asr

Pricing — 1 offering(s)

Microsoft Azure

LLM Speech transcription

  • $0.006 / minute Current 2026-04-02 → present

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how MAI-Transcribe fits into a cost-aware routing setup

See how →

Capability profile

How MAI-Transcribe rates across core capability dimensions, with the task-level evidence behind each rating.

transcription accuracy Strong

Microsoft reports MAI-Transcribe-1 as best-in-class on the FLEURS benchmark (avg WER 3.86% across 25 languages), beating Whisper-large-v3 on 25/25 languages, Gemini 3.1 Flash on 22/25, and GPT-Transcribe and ElevenLabs Scribe v2 on 15/25 each. Independent: the Artificial Analysis Speech-to-Text leaderboard lists MAI-Transcribe-1.5 at 2.4% WER and MAI-Transcribe-1 at 2.6% (AA's own eval, not FLEURS). Robust across accents and noisy real-world audio.

language support Moderate

43 languages in MAI-Transcribe-1.5 (25 in the deprecated MAI-Transcribe-1), with automatic language identification and a default multi-lingual mode. Fewer than the Azure Cognitive Services Speech-to-Text service's 100+ locales, though the covered set spans the major world languages plus a wide Indic range in 1.5.

speaker diarisation Weak

Not supported in either version. Microsoft lists diarization as planned but not yet available.

real time streaming Weak

File-based transcription only (WAV/MP3/FLAC, <300 MB per request). Real-time transcription is listed as out of scope / coming in a future release. Can be used as the transcription model inside the Voice Live API for voice-agent scenarios.

custom vocabulary Moderate

MAI-Transcribe-1.5 supports a phrase list (`phraseList`) for entity biasing on specialised terminology; MAI-Transcribe-1 has no customisation. No trainable custom model.

Benchmarks

Benchmark Score Config Source
WER (FLEURS, 25 languages — MAI-Transcribe-1, Microsoft-reported) 3.86 % Mean WER across MAI-Transcribe-1's 25 supported languages. Microsoft reports it as the lowest average against Scribe v2, Whisper-large-v3, GPT-Transcribe and Gemini 3.1 Flash on the same 25-language set. source ↗
WER (Artificial Analysis Speech-to-Text leaderboard — independent) — MAI-Transcribe-1.5 listed at 2.4% WER, MAI-Transcribe-1 at 2.6%, under Microsoft/Azure. AA's own evaluation set, not FLEURS — directionally consistent with Microsoft's best-in-class claim. source ↗

Operator guidance

Reach for MAI-Transcribe when you want strong published multilingual accuracy at a low price ($0.36/hour, ~2.8x cheaper than Azure Cognitive Services Speech-to-Text at $1.00/hour) and your workload is file-based. It is in public preview with no SLA, so it is not a fit for production workloads that need contractual guarantees — use microsoft-azure/speech-stt or Deepgram Nova-3 there. It also has no speaker diarization and no real-time streaming; route to speech-stt, Nova-3, or Amazon Transcribe for those. Its closest accuracy peers are OpenAI GPT-Transcribe and ElevenLabs Scribe v2. Only MAI-Transcribe-1.5 should be used — MAI-Transcribe-1 was deprecated on 20 Aug 2026.

Use cases

Limitations

Citations