Microsoft AI on Thursday introduced MAI-Transcribe-2, a speech-recognition model the company says is faster, more accurate, and cheaper than current offerings from OpenAI, Google, or ElevenLabs. The launch price is $0.10 per hour of audio.
Price and cost implications
The new price represents a substantial reduction compared with earlier models in this line: the first MAI Transcribe model, released five months ago, cost $0.36 per hour. The $0.10 launch price is roughly a 72% cut. For example, an enterprise that transcribes 100,000 hours of call-center audio per year would see costs fall from $36,000 to about $10,000 — a volume that is realistic for a large bank or telecom.
What MAI-Transcribe-2 offers and why it matters to enterprises
The model supports 60 languages, up from 43 in MAI-Transcribe-1.5 (June) and 25 in the original April release. It is available through Microsoft Foundry, the company’s model marketplace, and in the MAI Playground testing environment. Microsoft says MAI-Transcribe-2 was built for messy, real-world audio — background noise, low-fidelity recordings, and overlapping speech — rather than studio conditions.
More important than the raw language count is the set of features included in the base product:
- Speaker diarization: separates who spoke when in multi-party recordings.
- Word-level timestamps: attach precise timing to every word for search, editing, and syncing to video.
- Keyword biasing: allows developers to supply domain-specific vocabulary (drug names, product codes, employee names) so the model preserves jargon.
- Automatic language identification: no need to declare the language in advance.
Two additional capabilities are notable:
- Configurable output style: a "verbatim" mode that preserves fillers and false starts for compliance or legal use, and a "clean" mode that removes fillers for readable captions and notes.
- Code-switching support: handles conversations that switch languages mid-sentence (Microsoft cites Hinglish and Spanglish as examples), relevant for markets where a single support call may toggle languages repeatedly.
Specialty providers have historically charged premiums for many of these features; Microsoft includes them in the base offering at the launch price.
How to read Microsoft’s benchmark claims
Microsoft made three performance claims that rely on different benchmarks; technical buyers should understand what each measures and where limits lie:
-
FLEURS: Microsoft says MAI-Transcribe-2 ranks first on the FLEURS benchmark across 60 languages with an average word error rate (WER) of 5.2%. FLEURS, published in 2022, is built from native speakers reading about 2,000 sentences per language (roughly 12 hours per language). WER counts substitutions, insertions, and deletions versus a human reference; 5.2% means about one in twenty words is incorrect. FLEURS uses read speech, not conversational audio, and the average rose from 3.7% reported for MAI-Transcribe-1.5 in June — likely because expanding coverage from 43 to 60 languages folds in lower-resource languages where models typically perform worse. Buyers should request per-language breakdowns.
-
Artificial Analysis leaderboard: The model ranks second on Artificial Analysis’s WER leaderboard and, Microsoft says, defines that firm’s accuracy-latency Pareto frontier. Artificial Analysis tests models via their public APIs on a mix of simulated agent conversations, European Parliament speeches, and corporate earnings calls, with a heavy weighting toward English business speech. Being on the Pareto frontier means no rival is both faster and more accurate.
-
Raw speed: According to Artificial Analysis evaluations, MAI-Transcribe-2 is 10× faster than OpenAI’s GPT-Transcribe, 7× faster than ElevenLabs’ Scribe v2, and 5× faster than Google’s Gemini 3.5 Transcribe. In batch transcription, speed reduces GPU‑hour consumption and therefore cost: a model running at 300× real-time requires far fewer GPU hours than one running at 30×.
Caveats: FLEURS is read speech; WER does not capture diarization quality; and averages can hide per-language variation.
Three models in five months: the release cadence
The release pace is notable. MAI-Transcribe-1 launched April 2 with 25 languages at $0.36/hour. MAI-Transcribe-1.5 arrived June 2 with 43 languages, keyword biasing, and a third-place Artificial Analysis ranking. Now MAI-Transcribe-2 ships with 60 languages, diarization, timestamps, code switching, a second-place ranking, and a $0.10 price.
Three releases in five months, each increasing language coverage by roughly 40% and adding features competitors often reserve for higher tiers, suggests the team has converged on a stable architecture and is scaling data and compute — the phase when speech models typically improve rapidly. The organizational structure behind that speed, according to Mustafa Suleyman (Microsoft AI CEO), involved a small focused team freed from bureaucracy, supported by a larger group for vendor management and data acquisition.
Why Microsoft is building in-house models despite a $13 billion OpenAI investment
Microsoft has invested more than $13 billion in OpenAI and hosts OpenAI’s models across Azure, Office, and Copilot. Yet the company has moved toward greater independence. Since Mustafa Suleyman’s hire from Inflection AI in March 2024 and subsequent contract renegotiations with OpenAI, Microsoft has incrementally loosened exclusive arrangements and reduced revenue‑share obligations, enabling it to pursue its own MAI models and teams such as MAI Superintelligence.
A second driver is margin: routing prompts to third‑party models incurs costs. Routing to Microsoft’s own models on its own GPUs is cheaper. Bloomberg reported in July that Microsoft had begun answering some Word and Excel prompts with MAI models instead of third‑party models. Transcription is a natural early target because the problem is bounded and metrics are objective. Microsoft owns Teams (large volumes of meeting audio), Nuance (clinical documentation), and Azure Speech services — all candidate workloads to migrate to MAI‑Transcribe-2.
Mustafa Suleyman explicitly framed the shift as enabling product value at scale: build the capability once, deploy it across multiple products, and reduce payments to external partners that are becoming competitors.
Competitive picture: who is affected
Microsoft explicitly compares MAI-Transcribe-2 to OpenAI GPT-Transcribe, Google Gemini 3.5 Transcribe, OpenAI Whisper V3‑Large, and ElevenLabs Scribe v2. It does not call out specialist transcription providers such as Deepgram, AssemblyAI, Speechmatics, or Rev, even though those companies have sold enterprise transcription for years. The framing targets frontier labs: a focused model that matches accuracy while delivering 5–10× speed and a lower price can outcompete platform players.
Specialists’ remaining differentiation is domain depth — medical vocabularies, legal formatting, industry integrations — and Microsoft’s keyword-biasing feature appears aimed at narrowing that gap. Alibaba’s models remain a competitive benchmark on accuracy for parts of 2026; Microsoft does not claim to surpass Alibaba on accuracy.
Practical questions for technical decision-makers
The announcement leaves several operational questions to be resolved before switching vendors:
- Duration and pricing terms: Microsoft labels $0.10/hour a launch offer but provides no end date or standard rate — organizations should get those terms in writing.
- Streaming/real-time support: the release emphasizes batch throughput and long-form audio; it does not address real-time transcription for voice agents or live captions.
- Per-language accuracy: an average WER of 5.2% across 60 languages could mask significant variation; customers should test critical languages directly.
- Diarization error rates: WER does not measure speaker‑attribution errors; diarization quality should be validated separately.
- Data handling and privacy: the announcement does not specify data residency, retention, or whether Foundry submissions feed future training — key concerns for regulated industries.
These are routine follow-ups after a launch but are essential to move from leaderboard claims to production deployment.
What the strategy signals about Microsoft’s broader AI ambitions
Beyond speech, Microsoft’s MAI unit now ships models for images, voice, code, reasoning, and cybersecurity. At Build in June the company announced seven new MAI models in a single keynote. The playbook is consistent: target specific modalities, optimize inference costs, price below frontier labs, distribute through Foundry, and quietly replace third‑party models inside Microsoft products.
This is not an attempt to produce a single universal model that beats GPT or Gemini at everything. Rather, Microsoft appears to be assembling a portfolio of specialized models that together can serve most enterprise workloads internally and offer surplus capacity externally at prices that challenge specialists.
Availability
MAI-Transcribe-2 is available now through Microsoft Foundry and MAI Playground. The company is offering the $0.10 per hour introductory price, but did not specify how long that price will remain in effect.
