Microsoft MAI-Transcribe-2-Streaming Leads Benchmark

Microsoft AI has introduced MAI-Transcribe-2-Streaming, a real-time speech recognition model that has claimed the top spot on the Artificial Analysis streaming leaderboard. Evaluating 38 different models, the benchmark ranked the Microsoft system first with a 2.5% word error rate (WER) and a post-speech latency of just 0.13 seconds for final transcripts. This performance edged out Grok Voice Transcribe 2.0, which recorded a 2.7% WER. In the first partial transcript category, the Microsoft model also led with a 2.5% WER and 0.12 seconds of latency, outperforming ElevenLabs and Muse. By comparison, Grok Voice Transcribe 2.0 registered a 3.4% WER at 0.49 seconds, while Cartesia Ink-2 achieved a faster 0.07-second response but with a higher 4.0% WER.

The streaming model is priced at $0.54 per audio hour, or $9.00 per 1,000 minutes. This rate matches Gemini 3.5 Transcribe Live but sits above the pricing of ElevenLabs and Deepgram Flux. For non-streaming workloads, Microsoft is offering its batch model, MAI-Transcribe-2, at a promotional rate of $0.10 per audio hour through the end of 2026, while the older MAI-Transcribe-1.5 costs $0.36 per hour. At these rates, processing 1,000 audio hours costs $100 with the promotional batch model compared to $540 with the streaming version.

Developed by a core ten-person team with data and vendor support, the multilingual model covers 60 languages and supports diarization, word-level timestamps, and keyword biasing. It is currently available in public preview on Azure Speech in the East US, West US, Southeast Asia, and North Europe regions, as well as via Microsoft Foundry, MAI Playground, and OpenRouter. Live audio is processed through the Voice Live API, while prerecorded tasks use the Fast Transcription API. Because the live preview currently lacks a service-level agreement, Microsoft cautions against immediate production deployment without fallback systems.