Microsoft's new speech model is 10x faster and crazy cheap

Microsoft is officially picking a fight with Whisper, OpenAI, and ElevenLabs over speech-to-text dominance.

The company launched MAI-Transcribe-2, marketing it as the fastest, most accurate, and cheapest speech recognition model available today.

Crushing the latency curve

This is not just internal benchmark bragging. According to independent tests by Artificial Analysis, MAI-Transcribe-2 defines the current Pareto frontier for accuracy versus latency, ranking second on their overall Word-Error-Rate leaderboard.

When it comes to raw batch processing, the numbers are eye-opening. The model clocks in at 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs' Scribe v2, and 5x faster than Google's Gemini 3.5 Transcribe.

On the multilingual front, it topped the public FLEURS benchmark across 60 languages with an average Word-Error-Rate of just 5.2%.

Beyond raw speed, Microsoft packed the model with production-ready tooling that speech developers constantly struggle with.

It features native speaker diarization to separate distinct voices, alongside word-level timestamps for exact audio-to-text alignment and editing.

Developers can also feed in keyword biasing to prevent the model from mangling niche industry jargon, proper nouns, and acronyms.

Output formatting is split into two modes. A verbatim setting preserves every stutter and filler word for legal and compliance needs, while a clean setting strips the junk to deliver polished captions and meeting summaries.

It handles noisy environments, automatically detects languages, and supports code-switching for conversations naturally blending two tongues, like Spanglish or Hinglish.

MAI-Transcribe-2 is accessible right now on Microsoft Foundry, MAI Playground, and OpenRouter, with promotional pricing set at $0.10 per audio hour through the end of the year.

Why it matters

For builders running transcription pipelines at scale, this is an immediate headache-reliever. Cutting latency by up to 10x while paying just $0.10 an hour completely alters the unit economics of processing call logs, podcasts, and video archives.

Across the industry, this release puts massive margin pressure on specialized voice providers like ElevenLabs and cloud rivals like Google. Microsoft is using its sheer compute scale to turn high-quality transcription into a cheap, commoditized utility.

We think Microsoft played this launch smart by focusing on boring, essential developer features rather than flashy demos. Providing built-in diarization, keyword biasing, and verbatim toggles directly addresses enterprise pain points. The main caveat is that the aggressive ten-cent price tag is a limited-time promotional rate through the end of the year, so long-term customer retention will depend entirely on what the standard rate looks like in 2027.