On 3 September 2026, Microsoft AI launched MAI-Transcribe-2 with a blunt headline: the most capable and efficient speech-recognition model among its competitors. The announcement places it ahead, in Microsoft and Artificial Analysis evaluations cited, of Gemini 3.5 Transcribe, GPT-Transcribe, Whisper V3-Large and ElevenLabs’ Scribe v2.
On FLEURS, Microsoft reports first place across 60 languages with an average word-error rate of 5.2%. On Artificial Analysis, it says it defines the accuracy-latency Pareto frontier and ranks second on the word-error leaderboard, continuing gains from prior versions.
On speed, Artificial Analysis evals cited by Microsoft claim MAI-Transcribe-2 is about 10× faster than OpenAI’s GPT-Transcribe, 7× faster than Scribe v2 and 5× faster than Gemini 3.5 Transcribe, while delivering higher accuracy. Launch pricing is $0.10 per hour of audio as a limited-time offer through year-end.
The feature pack targets production: speaker diarization, word-level timestamps, domain keyword biasing, “verbatim” or “clean” styles, code switching (Hinglish, Spanglish), automatic language identification and robustness in noisy audio. One model across many languages reduces operational complexity.
Listed use cases include clinical note-taking, legal documentation, accessibility and captioning. The MAI-Transcribe-2-Streaming variant covers real-time transcription via API, while batch mode handles long files with lower relative latency than rivals, per Microsoft.
Availability: demo in Microsoft Foundry, MAI Playground and Open Router. Integration into Azure/Foundry Speech lets existing customers trial without changing cloud providers — an inertia argument against point APIs.
The launch fits MAI’s strategy to commoditize voice and image (alongside MAI-Image-2.6). In September–October 2026, the speech-to-text fight also featured xAI’s Grok Voice Transcribe 2.0, which answers with accuracy at the same price as its prior generation.
For enterprise buyers, the decision is no longer only “who has the lowest WER on a clean benchmark,” but latency on long audio, cost at scale, included diarization and real multilingual support. Microsoft packages those axes into a single leadership message.
Compliance owners should validate verbatim versus clean by sector: in health and legal, fillers and false starts are sometimes evidence; in public captions, they are noise. Keyword biasing helps with drugs, brands and acronyms.
Neural Pulse reports the speed, accuracy and price claims as Microsoft published them on 3 September; ModelDex and industry roundups amplified them. They do not replace a trial on the customer’s accents, noise and vocabulary.
The voice race is, at root, an agent race: without reliable transcription, a call-center agent or meeting copilot loses the thread. MAI-Transcribe-2 aims to be that cheap, fast layer under Microsoft assistants and third-party agents on Foundry.
October 2026 leaves a hyper-competitive speech market at $0.10/hour on several fronts. Who wins will not be only who publishes the best press release, but who sustains share and quality when the promotional offer ends.
