Skip to main content

MicrosoftAudio modelPreview

MAI-Transcribe-2

Popularityno data yet

n/ano search or usage data

Summary

MAI-Transcribe-2 is Microsoft’s file-oriented speech-to-text endpoint for 60 languages. Speaker attribution, timestamps, and domain biasing extend the spoken words with structured transcript information.

The API identifier is mai-transcribe-2. Foundry access is a public preview without an SLA, with an introductory price of $0.10 per hour of audio. Live input uses the separate streaming variant with its own identifier.

GraduallyBench Audio

The word error rate of MAI-Transcribe-2 is 3.4 % (35 errors over 1,040 reference words, German and English). The median of the 20 models with a German and English value is 4.5 %.

Word error rate, German and English

3.4 %

35 errors over 1,040 reference words

Text on silence

0 / 4

silence clips with text in the final transcript

Scope

36 / 36

valid of planned outputs

Access

OpenRouter

Word error rate, German and English, compared

Misrecognized words in percent (German and English), lower is better

20 of 20 models

As of 11/10/2026. Source: GraduallyBench Audio

MAI-Transcribe-2

Benchmark run
  • The route does not confirm the exact model version.
Show test details

Speech recognition

OpenRouter · German, English

36 valid outputs from 36 attempted and 36 planned test slots.

Word error rate (WER)
3.37 %(35 / 1,040 reference words)German, English
Character error rate (CER)
1.28 %(75 / 5,842 reference characters)German, English
Silence cases with words in the final transcript
0 %(0 / 4 assessable silence cases)German, EnglishAfter Unicode normalization, word segments containing letters or digits count. Punctuation-only output does not count.

WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 37 published benchmark results

Page 1 of 4

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Artificial Analysis AA-WER v2.0 (non-streaming) (Show measurement, test conditions, and source)
Source value
2
Metric
duration-weighted word error rate across AA-AgentTalk, VoxPopuli-Cleaned-AA, and Earnings22-Cleaned-AA
Unit
%
Benchmark version
AA-WER v2.0, current non-streaming board retrieved 2026-10-16
Category
audio
Direction
Lower is better
Source type
independent-evaluator
Evaluator
Artificial Analysis
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. Current board value for exact MAI-Transcribe-2 row. AA-WER v2 weights roughly eight hours of audio by duration: AA-AgentTalk 50%, VoxPopuli-Cleaned-AA 25%, Earnings22-Cleaned-AA 25%.
Context
Independent evaluator observation. This is distinct from Microsoft's vendor-reported 5.2% average FLEURS value in the prior record; AA-WER and FLEURS are different datasets and setups.
Source Artificial Analysis
Appen MedTerm-90 overall clinical entity recall (Show measurement, test conditions, and source)
Source value
81.31
Metric
Recall of clinician-validated medical entities
Unit
%
Benchmark version
MedTerm-90, September 2026 model versions
Category
audio
Direction
Higher is better
Sample
30017
Lower confidence bound
80.53
Upper confidence bound
82.09
Source type
independent-evaluator
Evaluator
Appen MedTerm-90
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. Appen reports the September 2026 Microsoft MAI-Transcribe 2 result on 1413 real physician dictations spanning 90.5 hours and 14 specialties. The reported recall is over 30017 clinician-validated entities; this entity total is the reported count, not the number of recordings. Microsoft's official API documentation identifies the batch transcription model as MAI-Transcribe-2; this result is not assigned to MAI-Transcribe-2-Streaming. Identical files, entities, thresholds, and judges were used for all models, without vocabulary hints. Appen reports a bootstrap 95% interval.
Context
This is clinical-entity recall, not WER or a general transcription score. The row is bound to MAI-Transcribe-2 batch transcription, not MAI-Transcribe-2-Streaming. Results represent September 2026 model versions under close-microphone single-speaker dictation conditions; Appen notes these conditions favor the models and may not generalize to other audio. The report page gives only an overall score, not the advertised per-specialty breakdown.
Source Appen MedTerm-90
PERIT-STT English WER, MAI-Transcribe-2 (Show measurement, test conditions, and source)
Source value
1.76
Metric
pooled normalized word error rate (WER)
Unit
%
Benchmark version
PERIT-STT English real-world leaderboard, captured 2026-10-17
Category
audio
Direction
Lower is better
Lower confidence bound
1.56
Upper confidence bound
1.99
Source type
independent-evaluator
Evaluator
PERIT-STT, Perit
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact PERIT-STT entry microsoft/mai-transcribe-2. English real-world speech-to-text leaderboard, pooled normalized word error rate against two independent human references with senior adjudication; 22 systems share the held-out audio. The method page states three runs per clip, but does not publish the clip count or how repeated runs are aggregated.
Context
The model-specific number of clips is not published, so no sample count is recorded. The method describes three runs per clip; repeat aggregation is unspecified. The 95% interval is reproduced as published.
Source PERIT-STT, Perit
Converse-STT 2.0 pooled WER, MAI-Transcribe 2 (Show measurement, test conditions, and source)
Source value
3.74
Metric
pooled normalized word error rate (WER)
Unit
%
Benchmark version
Ocular AI Converse-STT 2.0, technical report dated 2026-10-13
Category
audio
Direction
Lower is better
Sample
12555
Source type
independent-evaluator
Evaluator
Ocular AI, Converse-STT
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. OpenRouter model ID microsoft/mai-transcribe-2; MP3 and language=en with no additional request options. Four American English conversations, eight separate speaker tracks and 12,555 normalized reference words; pooled WER after Whisper English normalization. Batch transcription, no streaming or diarization.
Context
The evaluator does not name an immutable upstream serving revision for the OpenRouter route and publishes no confidence interval. 12,555 is the reference-word denominator, not a conversation or clip count.
Source Ocular AI, Converse-STT
DAI-ASR-I18N German WER, dual-channel (Show measurement, test conditions, and source)
Source value
6.3677542178
Metric
Normalized German word error rate (WER), dual-channel
Unit
%
Benchmark version
DAI-ASR-I18N v1.1, public per-language results embedded in page snapshot retrieved 2026-10-17
Category
speech recognition
Direction
Lower is better
Source type
independent-evaluator
Evaluator
David AI Research, DAI-ASR-I18N
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact model in DAI Table A1: Microsoft AI MAI-Transcribe-2 through Azure Speech. The evaluator used enhanced MAI-Transcribe-2 mode, verbatim transcription, word-level timestamps, diarization, and an explicit language locale; Azure Speech API version 2025-10-15. The captured page result row is model=mai_transcribe, metric=wer, channel=dual, language=de. DAI scores the two recorded channels separately and sums their edit counts. The public set provides about two evaluation hours per language across 21 languages; the German row's clip count and confidence interval are not published.
Context
The page does not publish a German clip count or confidence interval for this row. The article's cross-language aggregate figure is not substituted for this explicit German result.
Source David AI Research, DAI-ASR-I18N
Speko FLEURS Lithuanian lt_lt, batch WER (Show measurement, test conditions, and source)
Source value
6.1
Metric
Word error rate (WER), batch
Unit
%
Benchmark version
FLEURS lt_lt; Speko batch board dated 2026-10-02
Category
audio
Direction
Lower is better
Sample
50
Lower confidence bound
4.1
Upper confidence bound
8.3
Source type
independent-evaluator
Evaluator
Speko speech-to-text benchmarks
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. The Speko Lithuanian board identifies FLEURS lt_lt, n=50, corpus WER, batch, dated 2026-10-02. Its exact row azure:MAI-Transcribe-2 reports 6.1% WER and interval [4.1, 8.3].
Context
The per-language board metadata supplies n=50 for its FLEURS corpus, but the published interval does not specify its confidence level or calculation method. The exact MAI-Transcribe-2 row has a batch value; no MAI streaming value is present in that row.
Source Speko speech-to-text benchmarks
Speko FLEURS Indonesian id_id, batch WER (Show measurement, test conditions, and source)
Source value
2.1
Metric
Word error rate (WER), batch
Unit
%
Benchmark version
FLEURS id_id; Speko batch board dated 2026-10-09
Category
audio
Direction
Lower is better
Sample
50
Lower confidence bound
1.2
Upper confidence bound
3.4
Source type
independent-evaluator
Evaluator
Speko speech-to-text benchmarks
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. The Speko Indonesian board identifies FLEURS id_id, n=50, corpus WER, batch, dated 2026-10-09. Its exact row azure:MAI-Transcribe-2 reports 2.1% WER and interval [1.2, 3.4].
Context
The per-language board metadata supplies n=50 for its FLEURS corpus, but the published interval does not specify its confidence level or calculation method. The exact MAI-Transcribe-2 row has a batch value; no MAI streaming value is present in that row.
Source Speko speech-to-text benchmarks
Speko FLEURS Swahili sw_ke, batch WER (Show measurement, test conditions, and source)
Source value
9.6
Metric
Word error rate (WER), batch
Unit
%
Benchmark version
FLEURS sw_ke; Speko batch board dated 2026-10-09
Category
audio
Direction
Lower is better
Sample
50
Lower confidence bound
6.7
Upper confidence bound
12.5
Source type
independent-evaluator
Evaluator
Speko speech-to-text benchmarks
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. The Speko Swahili board identifies FLEURS sw_ke, n=50, corpus WER, batch, dated 2026-10-09. Its exact row azure:MAI-Transcribe-2 reports 9.6% WER and interval [6.7, 12.5].
Context
The per-language board metadata supplies n=50 for its FLEURS corpus, but the published interval does not specify its confidence level or calculation method. The exact MAI-Transcribe-2 row has a batch value; no MAI streaming value is present in that row.
Source Speko speech-to-text benchmarks
Omi batch medical WER, MAI-Transcribe-2 (Show measurement, test conditions, and source)
Source value
1.18
Metric
medical word error rate (Medical WER)
Unit
%
Benchmark version
Omi Medical Speech-to-Text Benchmark, September 2026, edition 2026-09-26.1
Category
audio
Direction
Lower is better
Sample
2872
Source type
community-reported
Evaluator
Omi Health Medical Speech-to-Text Benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Exact evaluated model row Microsoft MAI-Transcribe-2, batch mode. Omi's September 2026 edition evaluates 1,513 English recordings (7.2 hours). Medical WER uses 2,872 medical reference words, drug-term error rate 442 reference occurrences, and dose/unit matching 94 reference events. Timing covers 200 short clips under 30 seconds with three passes each (600 timing observations); third-party API timings were measured from Frankfurt. No immutable serving revision is named.
Context
Omi operates its own competing medical ASR product, Omi medical-1, and publishes this comparative evaluation itself; classify the results as community-reported, not as a neutral independent reproduction. Omi uses the same recordings and scorer for all compared models and states that small differences do not establish statistical significance. No confidence interval is published. The overall WER reference-word denominator is not reported; 1,513 is the shared recording count. Batch timing n=600 represents three passes over 200 clips, not 600 unique clips. The serving revision is not pinned.
Source Omi Health Medical Speech-to-Text Benchmark
Omi batch overall WER, MAI-Transcribe-2 (Show measurement, test conditions, and source)
Source value
5.64
Metric
overall word error rate (WER)
Unit
%
Benchmark version
Omi Medical Speech-to-Text Benchmark, September 2026, edition 2026-09-26.1
Category
audio
Direction
Lower is better
Source type
community-reported
Evaluator
Omi Health Medical Speech-to-Text Benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Exact evaluated model row Microsoft MAI-Transcribe-2, batch mode. Omi's September 2026 edition evaluates 1,513 English recordings (7.2 hours). Medical WER uses 2,872 medical reference words, drug-term error rate 442 reference occurrences, and dose/unit matching 94 reference events. Timing covers 200 short clips under 30 seconds with three passes each (600 timing observations); third-party API timings were measured from Frankfurt. No immutable serving revision is named.
Context
Omi operates its own competing medical ASR product, Omi medical-1, and publishes this comparative evaluation itself; classify the results as community-reported, not as a neutral independent reproduction. Omi uses the same recordings and scorer for all compared models and states that small differences do not establish statistical significance. No confidence interval is published. The overall WER reference-word denominator is not reported; 1,513 is the shared recording count. Batch timing n=600 represents three passes over 200 clips, not 600 unique clips. The serving revision is not pinned.
Source Omi Health Medical Speech-to-Text Benchmark
Omi batch drug-term error rate, MAI-Transcribe-2 (Show measurement, test conditions, and source)
Source value
2.04
Metric
drug-term error rate
Unit
%
Benchmark version
Omi Medical Speech-to-Text Benchmark, September 2026, edition 2026-09-26.1
Category
audio
Direction
Lower is better
Sample
442
Source type
community-reported
Evaluator
Omi Health Medical Speech-to-Text Benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Exact evaluated model row Microsoft MAI-Transcribe-2, batch mode. Omi's September 2026 edition evaluates 1,513 English recordings (7.2 hours). Medical WER uses 2,872 medical reference words, drug-term error rate 442 reference occurrences, and dose/unit matching 94 reference events. Timing covers 200 short clips under 30 seconds with three passes each (600 timing observations); third-party API timings were measured from Frankfurt. No immutable serving revision is named.
Context
Omi operates its own competing medical ASR product, Omi medical-1, and publishes this comparative evaluation itself; classify the results as community-reported, not as a neutral independent reproduction. Omi uses the same recordings and scorer for all compared models and states that small differences do not establish statistical significance. No confidence interval is published. The overall WER reference-word denominator is not reported; 1,513 is the shared recording count. Batch timing n=600 represents three passes over 200 clips, not 600 unique clips. The serving revision is not pinned.
Source Omi Health Medical Speech-to-Text Benchmark
Omi batch dose/unit matches, MAI-Transcribe-2 (Show measurement, test conditions, and source)
Source value
85.11
Metric
dose/unit reference-event match rate
Unit
%
Benchmark version
Omi Medical Speech-to-Text Benchmark, September 2026, edition 2026-09-26.1
Category
audio
Direction
Higher is better
Sample
94
Source type
community-reported
Evaluator
Omi Health Medical Speech-to-Text Benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Exact evaluated model row Microsoft MAI-Transcribe-2, batch mode. Omi's September 2026 edition evaluates 1,513 English recordings (7.2 hours). Medical WER uses 2,872 medical reference words, drug-term error rate 442 reference occurrences, and dose/unit matching 94 reference events. Timing covers 200 short clips under 30 seconds with three passes each (600 timing observations); third-party API timings were measured from Frankfurt. No immutable serving revision is named.
Context
Omi operates its own competing medical ASR product, Omi medical-1, and publishes this comparative evaluation itself; classify the results as community-reported, not as a neutral independent reproduction. Omi uses the same recordings and scorer for all compared models and states that small differences do not establish statistical significance. No confidence interval is published. The overall WER reference-word denominator is not reported; 1,513 is the shared recording count. Batch timing n=600 represents three passes over 200 clips, not 600 unique clips. The serving revision is not pinned.
Source Omi Health Medical Speech-to-Text Benchmark

Page 1 of 4

All specifications

Architecture

Model class
Speech recognition
Source: Model class

Input and output

Input
Recorded audio
Source: Input
Output
Transcript with timestamps and speaker attribution
Source: Output
Languages
60 languages
Source: Languages

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
Microsoft Foundry API
Source: Access
Shutdown
None announced
no source

API

API model ID
mai-transcribe-2
Source: API model ID

Further details

Preview terms
Public preview without an SLA
Source: Preview terms
Features
Speaker diarization, timestamps, and domain biasing
Source: Features

Prices

Introductory price
$0.10 per hour of audio as an introductory priceSource

Sources and data date

Show all 15 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

EvaluatorTest seriesComposite indicesRecorded active measurements
Artificial Analysis001

Without vendor claims and historical or disputed results; those stay labelled in the single values. 1 measurements are not yet assigned to a test series.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 14 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published

ocular-upstream-serving-revision-and-ci

Note

The Ocular report names the OpenRouter ID but no immutable upstream serving revision or WER confidence interval.

Additional sourceMicrosoft MAI Transcribe preview documentation (retrieved October 3, 2026; October 16, 2026; October 17, 2026) · Microsoft MAI Transcribe preview documentation · Editorial description reviewed October 3, 2026
Additional sourceMicrosoft MAI-Transcribe-2 release (retrieved October 3, 2026; October 4, 2026; October 16, 2026) · Microsoft MAI-Transcribe-2 release · Editorial description reviewed October 3, 2026
Additional sourceMicrosoft MAI-Transcribe-2 model page (retrieved October 3, 2026) · Microsoft MAI-Transcribe-2 model page · Editorial description reviewed October 3, 2026
Additional sourceArtificial Analysis (retrieved October 16, 2026)
Additional sourceAppen MedTerm-90 (retrieved October 16, 2026)
Additional sourcePERIT-STT, Perit (retrieved October 17, 2026)
Additional sourceOcular AI, Converse-STT (retrieved October 17, 2026)
Additional sourceDavid AI Research, DAI-ASR-I18N (retrieved October 17, 2026; October 19, 2026; October 20, 2026)
Additional sourceSpeko speech-to-text benchmarks (retrieved October 17, 2026)
Additional sourceOmi Health Medical Speech-to-Text Benchmark (retrieved October 17, 2026; October 23, 2026)
Additional sourceMicrosoft Foundry Labs (retrieved October 23, 2026)
Additional sourceModulate speech-to-text benchmark (retrieved October 23, 2026)
Additional sourceModel metadata (retrieved October 17, 2026)12
Additional sourceModel metadata (retrieved October 17, 2026; October 23, 2026)