MAI-Transcribe-2 Reality Check: 2.0% AA-WER and 410.7× Are Independent; $0.10/hr Is Temporary

What Microsoft actually released

Microsoft AI released MAI-Transcribe-2 on September 3, 2026 as a public-preview speech-to-text model for recorded audio. It supports 60 languages and adds speaker diarization, word-level timestamps, automatic language identification, keyword biasing, code-switching, and configurable clean or verbatim transcript styles.

This is an automatic-speech-recognition system, not a general-purpose reasoning or coding model. Microsoft exposes it through Microsoft Foundry/Azure Speech, the MAI Playground, and OpenRouter. The current Microsoft model page describes one hour of audio being processed in about 10 seconds of model inference and lists a limited-time price of $0.10 per audio hour.

Primary sources:

• Microsoft AI launch post
• Microsoft AI model page
• Azure Speech MAI-Transcribe documentation
• Microsoft Foundry announcement

The strongest independent evidence is 2.0% AA-WER at 410.7× real time

Artificial Analysis currently lists MAI-Transcribe-2 at 2.0% AA-WER v2, a 410.7× median speed factor, and $1.67 per 1,000 minutes on its non-streaming speech-to-text leaderboard.

Those numbers are independent of Microsoft's own FLEURS evaluation, and the methodology matters. Artificial Analysis says AA-WER v2 is an audio-duration-weighted score across roughly eight hours of audio from three datasets:

• AA-AgentTalk: 50%
• VoxPopuli-Cleaned-AA: 25%
• Earnings22-Cleaned-AA: 25%

For speed, Artificial Analysis defines the speed factor as input-audio seconds transcribed per second of processing time. Its displayed speed values are medians from benchmark trials in the last seven days using 10-minute audio files, so they can move as fresh runs arrive and may not predict very short clips or every provider region.

Independent benchmark:

• Artificial Analysis non-streaming speech-to-text leaderboard and methodology

The current table also makes an important correction to Microsoft's launch headline. MAI-Transcribe-2 is not literally first in each separate Artificial Analysis category: it is second on AA-WER at 2.0%, while Deepgram Nova-3 is faster at the current snapshot and several services are cheaper per 1,000 minutes. What MAI-Transcribe-2 does occupy is a strong accuracy-versus-speed Pareto position: among the displayed systems, improving on its accuracy generally means giving up throughput, and improving on its throughput means accepting higher WER.

That is a more defensible interpretation than turning “fastest, most accurate and cheapest” into three universal category wins.

Do not compare its 2.0% batch WER with streaming WER as if they were the same benchmark

Artificial Analysis places MAI-Transcribe-2 on its non-streaming benchmark. That matters because a batch system can inspect the uploaded recording rather than committing text while audio is still arriving.

For example, Meta's Muse Voice Transcribe is evaluated on Artificial Analysis's separate streaming benchmark, where latency is measured after detected end of speech. A 2.0% batch AA-WER and a streaming WER are therefore not directly rankable numbers. They use different serving modes, latency definitions, and product constraints.

Teams choosing a meeting archive or subtitle pipeline may care mostly about batch throughput and final accuracy. A live voice agent instead needs time-to-first/final transcript and endpoint behavior under streaming load. Those are different engineering questions.

Microsoft's multilingual FLEURS result is vendor-run and has two different averages

Microsoft reports that MAI-Transcribe-2 ranks first on FLEURS across its supported 60-language evaluation, with 5.2% average WER across 60 languages. Its model page separately highlights 3.4% average WER on the top 25 languages.

Those figures are not contradictory; they summarize different language subsets. But the distinction is important because quoting only the 3.4% value can make the broader 60-language result look stronger than it is.

FLEURS itself is a public multilingual speech benchmark derived from the FLORES sentence set. The public dataset covers 102 languages with parallel sentences and speaker-separated train versus development/test sets. Microsoft's exact MAI-Transcribe-2 scoring run is still vendor-reported: in the materials reviewed for this article, Microsoft does not publish a fully reproducible evaluation bundle with every per-language output, decoding setting, normalization rule, and raw prediction file needed to independently recreate its 5.2% and 3.4% aggregates.

Benchmark background:

• Google FLEURS dataset card

Confidence is therefore different by claim:

• High confidence: Artificial Analysis's current 2.0% AA-WER and 410.7× measurement under its documented non-streaming harness.
• Medium confidence: Microsoft's 5.2% 60-language and 3.4% top-25 FLEURS aggregates, because they are first-party evaluation results on a public benchmark but were not independently rerun here.
• Unproven as a universal claim: that the same accuracy holds across every accent, codec, noisy call, overlapping-speaker meeting, or domain vocabulary.

Pricing is compelling, but it expires

Microsoft's launch price is $0.10 per audio hour through December 31, 2026. Artificial Analysis normalizes that to about $1.67 per 1,000 minutes, which matches the arithmetic.

The word “limited-time” matters. Microsoft has not announced the standard post-promotion price in the reviewed launch material. A team projecting 2027 costs should therefore not build a long-term business case around $0.10/hour without a contracted price or updated rate card.

The price also should not be called the absolute cheapest speech recognition option from the Artificial Analysis table. Some lower-priced entries exist, though they differ substantially in accuracy, speed, hosting model, features, or provider terms. MAI-Transcribe-2's practical attraction is the joint combination of low WER, very high batch throughput, and a low promotional price.

Context is an audio-service limit, not an LLM token window

Microsoft does not publish a text-token context window, parameter count, downloadable checkpoint, or reasoning-effort control for MAI-Transcribe-2 in the materials reviewed here. Those concepts should not be inherited from unrelated MAI language models.

For this product, the meaningful limits are service constraints: accepted audio formats, file size, duration, region, concurrency, and preview availability. Microsoft Learn documents MAI-Transcribe through the Azure Speech fast-transcription path and describes input files under 300 MB on the MAI page.

There is an important documentation caveat: public Microsoft pages cited by users currently show different duration/file-size limits across related Speech references. Until Microsoft reconciles them, developers should treat the exact API version and endpoint documentation for their deployment as authoritative and test with representative long files.

Early preview feedback flags a diarization reliability issue on longer recordings

Two detailed Microsoft Q&A reports from September 4–5 are worth watching. They are not controlled benchmarks and do not establish a universal product defect, but they provide reproducible-looking practitioner evidence from the public preview.

One user reported native diarization succeeding through about 31 minutes and repeatedly returning HTTP 503 beyond roughly 33 minutes, while the same full recording transcribed successfully when diarization was disabled. A second user reported failures beginning around 15 minutes, with 408, 500, or 503 responses when diarization was enabled, again while substantially longer transcription succeeded without diarization.

Microsoft community responses did not identify a documented 15- or 32-minute diarization limit and recommended escalation to Azure support. Because the observed thresholds differ between users, the evidence is better described as an early preview reliability concern than as a fixed hard limit.

Public primary discussions:

• Microsoft Q&A: diarization failures above roughly 32 minutes
• Microsoft Q&A: diarization failures reported from roughly 15 minutes

The safest production takeaway is to test long multi-speaker files in the exact target region and API version before relying on native diarization. A preview service without a documented SLA should not be assumed to behave identically to a benchmark-only transcription request.

I did not recover a stable, substantive X post in this bounded review that added reproducible MAI-Transcribe-2 measurements beyond the official launch and the independent Artificial Analysis data, so no X quote or invented public consensus is included.

SWE-bench Verified and SWE-bench Pro are not applicable here

There is no MAI-Transcribe-2 SWE-bench Verified or SWE-bench Pro score in the evidence reviewed, and neither benchmark is appropriate for this speech-to-text model.

MAI-Transcribe-2 should not inherit coding, reasoning, tool-use, or agent benchmark scores from MAI-Thinking, MAI-Code, or any other Microsoft model. Its relevant evidence is ASR accuracy, multilingual robustness, diarization, timestamps, throughput, price, and service reliability.

This separation is especially important because Microsoft's model catalog contains other MAI systems with software-engineering benchmark claims. Similar naming does not make those results transferable.

Practical verdict

MAI-Transcribe-2 has unusually strong independent evidence for a new speech model: 2.0% AA-WER v2 at 410.7× median real-time throughput on Artificial Analysis's current non-streaming harness. Microsoft also offers a very aggressive $0.10/hour promotional price and a useful production-oriented feature set including diarization, timestamps, keyword biasing, code-switching, and 60-language support.

The limitations are just as important. The FLEURS multilingual results are vendor-run, the 3.4% top-25 and 5.2% 60-language averages describe different subsets, the promotional price expires at the end of 2026, and early public-preview users have reported native diarization failures on longer files. The model's batch benchmark should not be compared directly with streaming ASR scores.

For high-volume recorded-audio transcription, MAI-Transcribe-2 is one of the most compelling systems to test right now. For production adoption, the next evidence to demand is a version-pinned multilingual reproduction, long-form diarization reliability across regions, documented post-promotion pricing, and stable service-limit/SLA guidance.