The short version
On September 3, 2026, Microsoft released MAI-Transcribe-2, a speech-to-text model that transcribes 60 languages at $0.10 per hour of audio — a limited-time launch price — and processes an hour of audio in roughly ten seconds, which Microsoft says is 10x faster than OpenAI's GPT-Transcribe. The practical effect for engineering teams is that transcription is now effectively a commodity: accurate, multilingual, real-time and near-free. That opens features that used to be too slow or too expensive — live captions, call-center analytics, meeting notes, voice-driven search — but it also multiplies how much sensitive audio you capture and must govern. Treat it as an input to your AI and data strategy, not just a cheaper API line item.
What Microsoft launched on September 3
Microsoft's consumer-AI group released MAI-Transcribe-2, a speech-to-text model that it claims is the fastest, most accurate and cheapest in the world. It transcribes audio in 60 languages — up from 43 in June's interim release — and posts a 5.2% average word error rate across those languages, which puts it first on the multilingual FLEURS benchmark and second on Artificial Analysis' word-error-rate leaderboard. The model separates speakers, timestamps individual words, biases toward domain-specific keywords, handles conversations that switch between languages mid-sentence, and can output either verbatim or cleaned-up text. For teams building AI and data features, that feature set covers most real transcription needs out of the box.
Two numbers define the announcement. On speed, MAI-Transcribe-2 transcribes an hour of audio in about ten seconds — which Microsoft frames as roughly ten times faster than OpenAI's GPT-Transcribe, seven times faster than ElevenLabs' Scribe v2, and five times faster than Google's Gemini 3.5 Transcribe. On price, it costs $0.10 per hour of audio, a cut of more than 70% from the $0.36 that the earlier MAI-Transcribe-1 charged. Access runs through Microsoft Foundry using Azure Speech, the MAI Playground, and OpenRouter, so most teams can reach it through infrastructure they already use.
One caveat belongs next to the price, because it is the part that most affects planning. The $0.10 rate is explicitly a limited-time offer that Microsoft has said runs through the end of 2026, and the company has not published what the permanent price will be. Aggressive launch pricing is a familiar pattern in AI — a loss-leader to win developers — so the number to design your unit economics around is the eventual standard price, not the promotional one. The direction of travel is clear regardless: transcription is getting cheaper and faster across every vendor, not just this one.
Why price and speed change the calculus
When a capability gets an order of magnitude cheaper and faster at once, the interesting change is not that existing features cost less — it is that features that were previously uneconomic become viable. At $0.36 an hour and minutes of latency, transcribing every support call, every sales conversation, or every user voice note was a line item you had to justify and a delay users noticed. At $0.10 an hour and ten seconds per hour of audio, transcribing everything, in near-real-time, becomes the default assumption. A million hours of audio — a serious volume for a mid-sized product — transcribes for roughly $100,000, and likely less as prices keep falling. Cost stops being the reason not to build the feature.
Speed matters as much as price, because latency is what decides whether transcription can sit inside a live experience or only run as a batch job afterward. Ten seconds for an hour of audio means streaming transcription can keep up with a conversation, which is what live captions, real-time call-center guidance, meeting assistants, and voice-driven interfaces all require. The gap between “we transcribe the recording overnight” and “we show text as the person speaks” is the gap between an analytics feature and a product experience — and that gap just narrowed for everyone.
The strategic read is that speech-to-text is commoditizing the way object storage or push notifications did. When a capability becomes cheap, fast and widely available from several vendors, competitive advantage stops coming from having transcription and starts coming from what you do with it: how well you structure the transcript, extract intent, route it to the right workflow, and respect the user's privacy while doing so. The model is table stakes; the product around it is where the work — and the differentiation — now lives.
What it means for US & EU teams
The first implication is a build-versus-buy reversal for most teams. Self-hosting an open transcription model like Whisper on your own GPUs was often justified on cost when hosted APIs were pricey. At ten cents an hour with ten-second turnaround, that math flips: the engineering time to run, scale and maintain a transcription cluster rarely beats a managed endpoint. Keep self-hosting only where you genuinely need it — data that cannot leave your environment, offline or air-gapped operation, or heavily specialized vocabulary that demands a fine-tuned model. Otherwise, spend your engineers on the product, not on operating the model.
The second implication is that cheap transcription multiplies your sensitive-data footprint, and that is a governance problem, not just an architecture one. Voice recordings and their transcripts are personal data — frequently special-category data — under GDPR, and making them cheap to collect does not make them cheap to hold. You still need a lawful basis, consent where it applies, defined retention windows, and clear control over where audio is processed and stored. For regulated conversations in HealthTech or finance, treat the voice stream and its transcript as regulated data from the first prototype: decide what you record, what you keep, and for how long, before you turn the feature on.
The third implication is to keep your speech layer portable and jurisdiction-aware. Prices, accuracy and available languages are moving month to month across Microsoft, OpenAI, Google and ElevenLabs, so put transcription behind a provider-agnostic interface rather than wiring one vendor's SDK through your codebase — switching should be a config change, not a rewrite. That portability is also a data-residency lever: for EU users, confirm which region a speech API actually runs in and whether the provider is subject to foreign data-access law, and be ready to place inference where both price and jurisdiction line up. Cheap does not mean compliant by default.
A checklist before you add voice
- Model the permanent price. Base unit economics on the expected standard rate, not the $0.10 launch promotion; re-check as vendors reprice.
- Buy, don't host — usually. Default to a managed API; reserve self-hosting for data-residency, offline, or specialized-vocabulary cases where it clearly pays off.
- Abstract the provider. Put transcription behind an interface so you can follow price, accuracy and language coverage across vendors without a rewrite.
- Decide retention up front. Set explicit rules for how long you keep audio and transcripts, and delete on schedule — cheap capture is not a reason to hoard.
- Get consent and a lawful basis. Treat voice as personal (often sensitive) data under GDPR; collect consent where required and document your basis.
- Match region to jurisdiction. Confirm where the speech API processes audio for EU users and whether it is exposed to foreign data-access law.
- Test on your real audio. Benchmark accuracy on your accents, noise and domain terms — and use keyword biasing before assuming a benchmark score transfers.
Frequently asked questions
What is Microsoft MAI-Transcribe-2?
MAI-Transcribe-2 is a speech-to-text model Microsoft launched on September 3, 2026. It transcribes audio in 60 languages, tops the FLEURS multilingual benchmark with a 5.2% average word error rate, and processes an hour of audio in about ten seconds. Microsoft describes it as 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs' Scribe v2, and 5x faster than Google's Gemini 3.5 Transcribe. It supports speaker diarization, word-level timestamps, keyword biasing, code-switching between languages, and verbatim or clean output, and is available through Microsoft Foundry using Azure Speech, the MAI Playground, and OpenRouter.
How much does MAI-Transcribe-2 cost?
Microsoft priced MAI-Transcribe-2 at $0.10 per hour of audio, which it describes as a limited-time offer through the end of 2026, and has not published a permanent price. That is roughly a 72% cut from the $0.36 per hour of the earlier MAI-Transcribe-1. At $0.10 an hour, transcribing a million hours of audio costs about $100,000, so for most products the cost of turning speech into text is no longer a meaningful blocker. Model the eventual standard price rather than building a business case that only works at the launch rate.
Should teams self-host Whisper or use a hosted speech API?
For most teams, a hosted API now wins on total cost. When a managed model transcribes an hour of audio for ten cents in ten seconds, self-hosting an open model like Whisper on your own GPUs rarely pays off once you count engineering time, GPU rental, scaling and maintenance. Self-hosting still makes sense when data cannot leave your environment for regulatory reasons, when you need offline or air-gapped operation, or when you have highly specialized vocabulary that requires a fine-tuned model. Otherwise, spend the engineering effort on the product around transcription, not on running the model.
What are the compliance risks of adding transcription?
Voice recordings and their transcripts are personal data, and often sensitive data, under GDPR. Adding cheap transcription does not remove the obligations: you still need a lawful basis, consent where required, clear retention limits, and control over where audio is processed and stored. For EU users, check which region the speech API runs in and whether the provider is subject to foreign data-access laws. In HealthTech and FinTech, treat call recordings and voice notes as regulated data from the first prototype. Cheaper inference lowers the cost barrier but raises the volume of sensitive data you are responsible for.
Sources
Microsoft AI — MAI-Transcribe-2 is the fastest, most accurate and cheapest speech recognition model in the world (primary source, September 3, 2026)
VentureBeat — Microsoft AI's MAI-Transcribe-2 undercuts OpenAI, Google and ElevenLabs on price and speed (September 3, 2026)