MAI-Transcribe 2 transcription online

What is MAI-Transcribe 2?

MAI-Transcribe 2 is Microsoft's automatic speech recognition model, succeeding MAI-Transcribe 1.5. It produces a transcript, word-level timestamps, and — new in this version — word-level timestamps and cleaner transcript controls, and it can guide the model towards domain-specific vocabulary before a transcription run begins.

Microsoft reports MAI-Transcribe 2 improves on MAI-Transcribe 1.5 and ranks on the Artificial Analysis Word Error Rate (AA-WER) leaderboard and the FLEURS benchmark. We haven't published our own accuracy or speed measurement for this model yet, so treat vendor-published rankings as a starting point rather than a number you can reproduce exactly on your own audio.

Entity biasing sets it apart from most general-purpose ASR models: you can supply up to 200 words or phrases — names, product terms, medical terminology, legal concepts — that the model should expect to encounter. The decoder is biased towards these candidates when audio is phonetically ambiguous, which substantially reduces first-pass errors on proper nouns and specialist vocabulary. Biasing combines with diarization and either transcript style in the same request.

You can also choose how the transcript reads: Clean removes filler words and false starts for a readable result, while Verbatim preserves them — useful when the exact spoken record matters, such as legal or research work.

MAI-Transcribe 2 vs. Whisper vs. Gemini 3.5 Transcribe

All three models cost €0.02/min. Full comparison →

All three models' transcripts can also be translated afterward into any supported language using the dashboard's Translate action — the row above only covers Whisper's free, single-pass translation option at upload time.

MAI-Transcribe 2's row reflects Microsoft's own vendor-published claims, not a convert.express measurement; Whisper's comes from published clean-benchmark and real-world comparisons; Gemini 3.5 Transcribe's is a vendor-published average. Different benchmarks, different conditions — not a strict apples-to-apples comparison.

Frequently asked questions

What is MAI-Transcribe 2?

MAI-Transcribe 2 is Microsoft's automatic speech recognition model, the successor to MAI-Transcribe 1.5. It adds word-level timestamps, a choice of Clean or Verbatim transcript style, and supports up to 200 entity-biasing phrases — proper nouns, domain terms, or names that should be transcribed exactly. On convert.express, optional speaker labels are added in a separate whole-file pass using pyannote Precision-3.

Does MAI-Transcribe 2 support speaker labeling?

Yes. MAI-Transcribe 2 returns the transcript and word timestamps without native diarization. Alongside it, a separate pyannote Precision-3 job analyzes the whole normalized recording, and convert.express aligns its speaker time ranges to the MAI words. Running speaker detection once over the full file preserves a single speaker timeline even when very long audio needs multiple Azure transcription requests.

What is the difference between Clean and Verbatim transcript styles?

Clean removes filler words, false starts, and self-corrections for a readable transcript — the default. Verbatim preserves them, including repetitions and mid-sentence corrections, which is useful for legal, research, or linguistic work where the exact spoken record matters. You can pick either style when MAI-Transcribe 2 is selected, and it combines with speaker labeling and entity biasing.

What is entity biasing and how do I use it?

Entity biasing (also called vocabulary hints) lets you supply a list of words and phrases — up to 200 — that the model should expect to encounter. This primes the ASR decoder to prefer these candidates when the audio is ambiguous. Use it for: interviewee names, organisation names, product names, technical terminology, or any proper noun that a general-purpose model might mis-transcribe. Enter the phrases in the transcription settings before uploading your file.

How accurate is MAI-Transcribe 2 compared to Whisper?

Microsoft reports MAI-Transcribe 2 improves on MAI-Transcribe 1.5 and ranks on both the Artificial Analysis Word Error Rate (AA-WER) leaderboard and the FLEURS benchmark. We haven't published our own head-to-head measurement against Whisper yet, so treat vendor-published rankings as a starting point rather than a guarantee for your specific audio. In practice the difference is most visible on proper nouns and entity-heavy content once you've supplied biasing phrases. On low-resource or niche languages, Whisper may still be more reliable since it covers 57 languages vs. 60 for MAI.

What does "Public Preview" mean?

MAI-Transcribe 2 is in public preview on convert.express (as is Gemini 3.5 Transcribe), which means it is available but carries no uptime SLA. If you pick a model explicitly, that is the model your job runs on — convert.express never silently transcribes with a different one, so if your chosen model is unavailable the job fails and tells you why. Jobs left on the default model do fall back: Whisper runs first, and MAI-Transcribe 2 picks the job up if Whisper cannot take it. If your workflow depends on guaranteed availability, choose Whisper as the explicit model. Preview status also means the model may be updated at any time, which could change output format or accuracy characteristics.

When should I choose MAI-Transcribe 2 over Whisper or Gemini 3.5 Transcribe?

Choose MAI-Transcribe 2 when: you need word-level timestamps and optional whole-file speaker labels; you can provide up to 200 entity biasing phrases to improve accuracy on proper nouns; you want a choice between a readable (Clean) or exact (Verbatim) transcript; and the language you need is in its 60-language set. Choose Whisper when you need a language outside MAI's coverage, want a free single-pass English translation at upload time, or prefer a generally-available model with formal SLA guarantees. Choose Gemini 3.5 Transcribe when you need a much larger biasing vocabulary (up to 1,000 terms) or the widest language coverage of the three. Either way, once a job finishes you can translate its transcript into English or any other supported language from your dashboard — that option isn't limited to Whisper.