Felo API PlatformFelo API Platform
Back to models

Microsoft AI: MAI-Transcribe 2 API

MAI-Transcribe2 is a multilingual speech-to-text model for accurate transcription, with automatic language identification, speaker diarization, and word-level timestamps.

Developed in-house by the Microsoft AI team, MAI-Transcribe2 converts audio into text and supports60 languages with automatic language identification. It can handle code switching in mixed-language speech, identify speakers, add word-level timestamps, and use keyword biasing for specialized terminology. Transcription style can be configured as verbatim or clean.

The model is suited to captions, subtitling, accessibility, call transcription, and voice-enabled applications. Teams can use it to create meeting records with speaker attribution and searchable transcripts for summarization and action-item extraction. It also fits contact center and sales call review, recorded voice-agent analysis, and timestamp-aligned video localization. Microsoft reports that it transcribes one hour of audio in about10 seconds of inference; the model is a batch offering in public preview, while real-time transcription requires a separate Voice Live setup.

Reported benchmark results include a5.2% average word error rate on the FLEURS multilingual benchmark across60 languages, where it ranked number one. These are vendor-reported figures, and the benchmark table is Microsoft's own rather than an independent verdict.

Updated:

MODEL ID

mai-transcribe-2
Context
0
Max output
0
Input
audio
Output
transcription

Highlights

  • 60 languages
  • Automatic language identification
  • Mixed-language code switching
  • Speaker diarization
  • Word-level timestamps
  • Verbatim or clean styles

Providers

1 providers

Compare API prices across providers. Select a provider to view all charges and conditions.

Provider · Pricing details
Uptime
$0.000027777777777777777777777777777777777777777777777778

Benchmarks

  • FLEURS multilingual benchmark (60 languages)

    Average word error rate; ranked number one.

    5.2%
  • FLEURS multilingual benchmark (top25 languages)

    Word error rate.

    3.4%
  • Artificial Analysis word-error-rate leaderboard

    Approximate error rate; ranked second.

    2.5%
  • DAI-ASR-I18N

    Had the lowest transcription error rate across the tested languages.

    18 of21
  • DAI-ASR-I18N

    Languages had transcription error below10%.

    16

Scores as published by the model provider.

API reference

Model-specific protocol details are not published. Check the API reference for integration requirements.

API reference →

Frequently asked questions

What is Microsoft AI: MAI-Transcribe 2?

Developed in-house by the Microsoft AI team, MAI-Transcribe2 converts audio into text and supports60 languages with automatic language identification. It can handle code switching in mixed-language speech, identify speakers, add word-level timestamps, and use keyword biasing for specialized terminology. Transcription style can be configured as verbatim or clean. The model is suited to captions, subtitling, accessibility, call transcription, and voice-enabled applications. Teams can use it to create meeting records with speaker attribution and searchable transcripts for summarization and action-item extraction. It also fits contact center and sales call review, recorded voice-agent analysis, and timestamp-aligned video localization. Microsoft reports that it transcribes one hour of audio in about10 seconds of inference; the model is a batch offering in public preview, while real-time transcription requires a separate Voice Live setup. Reported benchmark results include a5.2% average word error rate on the FLEURS multilingual benchmark across60 languages, where it ranked number one. These are vendor-reported figures, and the benchmark table is Microsoft's own rather than an independent verdict.

How is usage billed?

Your bill depends on the provider you use, its prices and your usage. Longer inputs or other conditions may have different prices.

What input and output does it support?

It accepts audio input and returns transcribed text.

How many languages does it support?

It supports60 languages with automatic language identification and can handle code switching in mixed-language speech.

What tasks is it suited to?

It is suited to captions, call transcription, subtitling, accessibility, meeting records, and voice-enabled applications. Speaker attribution and word-level timestamps can support searchable records and aligned subtitles.

What are its context and output limits?

The provided information does not specify a context window or output limit.

How does it differ from sibling models?

The provided information identifies MAI-Transcribe1.5 and MAI-Transcribe-2-Streaming as siblings but does not describe a direct comparison. It notes that the streaming variant delivers first partial transcripts in about100 milliseconds.