Microsoft MAI-Transcribe-2 leads benchmark

The real-time speech model placed first on the Artificial Analysis leaderboard with a 2.5% error rate, but promotional claims on X triggered an immediate correction from ElevenLabs.

Microsoft AI launched MAI-Transcribe-2-Streaming on October 1, 2026, targeting autonomous agents and real-time voice applications. The speech recognition model debuted at the top of the independent Artificial Analysis benchmark, posting a 2.5% word error rate (the share of misplaced or dropped words) and 0.13 seconds of final transcript latency. In the same official announcement, division chief Mustafa Suleyman introduced speech synthesis engines MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The strong technical debut, however, collided with a public messaging blunder: speed and cost comparisons Suleyman shared on X contradicted actual market pricing.

Technical lead for MAI-Transcribe-2

In the Artificial Analysis evaluation covering 38 competing speech models, MAI-Transcribe-2-Streaming edged out direct competitors in accuracy and response time. Microsoft finished ahead of Grok Voice Transcribe 2.0 (2.7% error rate and 0.49-second latency), Meta’s Muse Voice Transcribe (3.1% and 0.16 seconds), Cartesia Ink Preview (3.1% and 0.11 seconds), and ElevenLabs Scribe v2 Realtime (3.6% and 0.14 seconds).

According to technical documentation, the model supports 60 languages with automatic language identification. It is engineered for the intake stage of conversational agents. In these pipelines, partial transcriptions must arrive without perceptible delay so an agent can process context before a person stops speaking, preventing awkward gaps in dialogue.

The pricing dispute with ElevenLabs

The dispute arose when Suleyman posted on X that the new model ran 55% faster and 60% cheaper than ElevenLabs. Mati Staniszewski, chief executive of ElevenLabs, responded the following day on X: “Quick note back: Scribe v2.5 streaming is actually 28% cheaper per hour than MAI-Transcribe-2, not 60% more expensive. And a new version shipping soon :)”.

Public pricing confirms Staniszewski’s correction. Microsoft set an introductory streaming price of $0.54 per audio hour through the end of 2026, which works out to $9.00 per 1,000 minutes. ElevenLabs charges $6.50 per 1,000 minutes for streaming audio. In practice, Microsoft’s streaming service runs roughly 38% higher than its rival. For asynchronous batch processing, Microsoft charges $0.10 per hour, or $1.67 per 1,000 minutes.

The discrepancy stems from an internal messaging slip. Microsoft’s product announcement credits the 55% speed increase and 60% cost savings to its speech generator, MAI-Voice-2.1-Flash, when compared with competing audio generation models. Suleyman transposed those text-to-speech figures onto the transcription launch post.

Building an independent voice stack

The release signals Suleyman’s effort to assemble proprietary core models inside Microsoft AI, reducing reliance on OpenAI endpoints. Alongside the transcription tool, Microsoft shipped MAI-Voice-2.1, which covers 23 languages and 26 regional accents while keeping vocal tone consistent across languages, and MAI-Voice-2.1-Flash, which generates 45 seconds of audio with an initial response latency of 150 milliseconds.

In conversational systems, the loop between listening, reasoning, and speaking needs to finish in under 500 milliseconds to feel lifelike. By packaging low-latency transcription and synthesis inside Azure and Microsoft Foundry, Microsoft wants enterprise voice developers building inside its own cloud environment.

The models run in public preview on Microsoft Foundry (with data centers in the United States, Sweden, and Southeast Asia), MAI Playground, and through integrations on Vercel, OpenRouter, and LiveKit. There are no open weights or local downloads available, as access is restricted to paid cloud APIs. RUMOR: Microsoft has not published any schedule for bringing these models to Windows or Microsoft 365 Copilot, and open releases remain unconfirmed.

Sources

• Microsoft AI official announcement: https://news.microsoft.com/source?p=25993
• Microsoft Learn technical documentation: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/mai-transcribe-2-streaming
• Mustafa Suleyman announcement post on X: https://x.com/mustafasuleyman/status/2105699115602677984
• Artificial Analysis / ExplainX benchmark report: https://explainx.ai/blog/microsoft-mai-transcribe-2-streaming-aa-wer-2026
• Mati Staniszewski reply on X: https://x.com/mati/status/2106055816172486671