MAI-Transcribe-2 - Speech Service

Learn how to use the MAI-Transcribe model via Azure Speech API.

This feature is currently in public preview. This preview is provided without a service-level agreement, and is not recommended for production workloads. Certain features might not be supported or might have constrained capabilities. For more information, see Supplemental Terms of Use for Microsoft Azure Previews.

MAI-Transcribe is a next‑generation speech-to-text model built in‑house by the Microsoft AI team. It delivers fast and accurate transcription across 60 languages and real-world audio conditions. MAI-Transcribe-2 supports a wide range of workloads including video captioning, meetings, clinical notes, call center documentation, accessibility tools, content creation, and voice agents. The model provides speaker diarization, strong performance in noisy environments, word-level timestamps, automatic language identification, keyword biasing, code switching, and configurable transcription styles (clean transcripts without fillers, or verbatim).

The following models are supported:

• MAI-Transcribe-2
• MAI-Transcribe-1.5
• MAI-Transcribe-1: Deprecated on Aug 20, 2026.

Prerequisites

• An Azure subscription. You can create one for free.
• A Microsoft Foundry resource for Speech in the Azure portal.
• The Speech resource key and region. After your Speech resource is deployed, select Go to resource to view and manage keys. For the current list of supported regions, see Speech service regions.
• An audio file in one of these formats: WAV, MP3, or FLAC. For the maximum file size and audio duration, see the audio parameter in the Transcriptions - Transcribe REST reference. If you enable speaker diarization, see the note on recording length in the next section.

Use a MAI-Transcribe model

Use MAI‑Transcribe‑2 to generate transcripts from audio input. Configure the following features through the corresponding API parameters:

Feature	Parameter	Values	Default	Description
Model selection (Required)	enhancedMode. model	"MAI-Transcribe-2"	—	Set to call the MAI‑Transcribe model.
Enhanced mode (Required)	enhancedMode. enabled	true | false	false	Set to true to call MAI-Transcribe.
Speaker diarization	diarization. enabled	true | false	false	Segments the recording by speaker and attributes each segment to the correct person. Returns speaker-labelled segments with speaker, offsetMilliseconds, and durationMilliseconds metadata. Supports shorter recordings than transcription without diarization; see the note that follows this table.
Word-level timestamps	modelOptions. timestamps	"word" | "segment" | "none"	none	Controls timing granularity. "word" returns offsetMilliseconds and durationMilliseconds for every word, enabling precise alignment, search, navigation, and editing. "segment" returns timing information for each segment: the output transcript is partitioned into segments by language and speaker. So when diarization is enabled, a segment is created for each speaker and language, and when it’s disabled it’s created for each language. "none" omits timing data.
Keyword biasing	phraseList.phrases	Array of strings	[]	Biases recognition toward supplied keywords. Use for domain-specific terminology, abbreviations, product names, and proper nouns that are hard to disambiguate from context alone. Terms are hints, not forced output.
Transcription styles	modelOptions. transcribeStyle	"verbatim" | "clean"	"verbatim"	Controls output style. "verbatim" captures speech exactly as spoken, including filler words and false starts, for compliance, QA, and analysis workloads. clean removes fillers to produce readable captions, notes, and published transcripts.
Language selection	locales	["language_code"]	Unspecified (Automatic Detection)	Allows forcing transcription in a specific language. This is a very strong hint to the model, and don’t specify it unless you’re absolutely certain that the recording is in the given language, and the default auto-detection doesn’t work. Only one language can be provided here. MAI-Transcribe-2 supports 60 languages.
Code switching	Automatic	—	—	For commonly blended language pairs such as Hinglish and Spanglish, handles conversations that move between languages mid-utterance.
Noise robustness	Inherent	—	—	Maintains transcription quality on audio recorded outside controlled environments, including background noise, overlapping speech, and variable microphone quality.

Language support

By default, the model operates in multilingual mode. The following languages are currently supported:

Language code	Language	MAI-Transcribe-1.5 support	MAI-Transcribe-2 support
af	Afrikaans		✅
ar	Arabic	✅	✅
as	Assamese	✅	✅
az	Azerbaijani		✅
bg	Bulgarian	✅	✅
bn	Bengali	✅	✅
bs	Bosnian		✅
ca	Catalan	✅	✅
cs	Czech	✅	✅
da	Danish	✅	✅
de	German	✅	✅
el	Greek	✅	✅
en	English	✅	✅
es	Spanish	✅	✅
et	Estonian	✅	✅
fa	Persian		✅
fi	Finnish	✅	✅
fil	Filipino		✅
fr	French	✅	✅
gl	Galician		✅
gu	Gujarati	✅	✅
he	Hebrew		✅
hi	Hindi	✅	✅
hu	Hungarian	✅	✅
hy	Armenian		✅
id	Indonesian	✅	✅
is	Icelandic		✅
it	Italian	✅	✅
ja	Japanese	✅	✅
kk	Kazakh		✅
kn	Kannada	✅	✅
ko	Korean	✅	✅
lt	Lithuanian	✅	✅
lv	Latvian		✅
mk	Macedonian		✅
ml	Malayalam	✅	✅
mr	Marathi	✅	✅
ms	Malay		✅
nb	Norwegian Bokmål	✅	✅
ne	Nepali		✅
nl	Dutch	✅	✅
or	Odia	✅	✅
pa	Punjabi (Gurmukhi script)	✅	✅
pl	Polish	✅	✅
pt	Portuguese	✅	✅
ro	Romanian	✅	✅
ru	Russian	✅	✅
sk	Slovak	✅	✅
sl	Slovenian	✅	✅
sv	Swedish	✅	✅
sw	Swahili		✅
ta	Tamil	✅	✅
te	Telugu	✅	✅
th	Thai	✅	✅
tr	Turkish	✅	✅
uk	Ukrainian	✅	✅
ur	Urdu		✅
vi	Vietnamese	✅	✅
yue	Cantonese		✅
zh	Chinese (simplified)	✅	✅

Use MAI-Transcribe with Voice Live

You can also use the MAI-Transcribe model for input audio transcription in the Voice Live API. Set the model field in the input_audio_transcription session configuration. For details, see How to customize Voice Live input and output.

Related content

• For more information about using LLM Speech API, see LLM Speech API
• MAI-Voice in Azure Speech
• How to customize Voice Live input and output