Microsoft's new streaming transcription model tops Artificial Analysis rankings

Microsoft AI has released MAI-Transcribe-2-Streaming, its first real-time speech-to-text model, and the testing firm Artificial Analysis ranks it first out of 38 models for transcription accuracy. The model launched October 1, 2026, alongside two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. It is built for voice agents, live captioning and dictation tools, where how fast a system turns speech into usable text shapes the whole experience.

The model is the real-time counterpart to MAI-Transcribe-2, a batch transcription model Microsoft released in September that processes complete audio files rather than live streams. MAI-Transcribe-2-Streaming instead takes audio in continuously and sends text back while a person is still speaking, covering 60 languages with automatic, ongoing language detection that does not require a speaker to specify which language they are using.

According to Artificial Analysis benchmarks cited by Microsoft, the model produces its first draft of a transcript, called a partial, about 100 milliseconds after receiving audio, then revises that draft as more context arrives before locking in a final transcript. On the firm’s AA-WER Streaming index, built from roughly eight hours of audio drawn from agent conversations, European parliament speeches and company earnings calls, the final transcript reached a 2.5% word error rate (the share of words transcribed wrong) just 0.13 seconds after a speaker stopped talking. Its first partial scored the same 2.5% error rate at 0.12 seconds, meaning the earliest guess was already as accurate as the polished version. Microsoft says its internal testing shows words appearing twice as fast as its nearest competitor.

Two rivals placed close behind on the same index: xAI’s Grok Voice Transcribe 2.0 scored 2.7% error at 0.49 seconds, and Meta’s Muse Voice Transcribe scored 3.1% error at 0.16 seconds. One competitor, Cartesia’s Ink-2, returned final transcripts faster, at 0.07 seconds, but with a higher 4.0% error rate, illustrating the speed-versus-accuracy trade-off that Microsoft says its model avoids.

The speed comes at a premium. Streaming transcription costs $0.54 per hour of audio, an introductory rate Microsoft is running through the end of 2026, compared with $0.10 per hour for the batch version. Artificial Analysis converts that to $9 per 1,000 minutes, pricing it above xAI’s and Meta’s streaming rates and roughly in line with Google’s.

Microsoft offers two ways to connect to the model: a WebSocket-based Realtime API compatible with existing OpenAI Realtime integrations, and the Azure Speech SDK, which handles connection management, retries and audio streaming directly. The model is also available through the MAI Playground, Vercel and Azure Voice Live, with support for LiveKit, a platform for real-time audio and video apps, listed as coming soon.

The release remains a public preview. Microsoft has not published a service-level agreement, and the model’s weights are not open. Whether it can identify which speaker said what in a live, multi-person stream is not yet stated, and that gap will likely matter most to developers building voice agents meant to handle more than one speaker at a time.