MAI-Transcribe-2 Integrates Speaker Diarization and Word-Level Timestamps Into a Single Speech API, With Batch Processing at 410× Real Time

Microsoft has released MAI-Transcribe-2, bringing speaker diarization, word-level timestamps, and output-style controls—capabilities that previously often required separate post-processing services—into a single speech-to-text model. The API can automatically detect languages, use keyword biasing to improve recognition of proper nouns, and switch between “verbatim” transcription and a “clean” mode that removes filler words. Microsoft also specifically highlights support for in-sentence code-switching such as Hinglish and Spanglish, as well as noisy audio, overlapping speech, and long recordings.

Microsoft claims the model achieves an average word error rate of 5.2% across the 60 languages in FLEURS, but it has not published per-language results. This is particularly important for Chinese-language deployments: a cross-language average may obscure differences involving Traditional Chinese, accents, specialized terminology, and mixed Chinese-English speech. FLEURS also consists primarily of read speech, so it is not representative of customer-service calls or multi-speaker meetings.

More readily verifiable are Artificial Analysis’s non-streaming tests conducted through public APIs. Its current data shows that MAI-Transcribe-2 records an AA-WER of 2.0% and processes approximately 410.7 seconds of audio per second. GPT-Transcribe posts 3.3% and 40.0× real-time performance, Scribe v2 achieves 2.2% and 53.6×, while Gemini 3.5 Transcribe reaches 2.6% and 89.9×. These results place the new model on the Pareto frontier for accuracy and speed. Nova-3 remains faster at 606.1× real time, though with a higher error rate.

The immediate engineering impact is that batch subtitling, recording archival, and speech-data annotation can be completed with less waiting, while reducing error propagation among diarization, punctuation, and time-alignment pipelines. The model is available through Microsoft Foundry, MAI Playground, and OpenRouter, with promotional pricing of $0.10 per hour through the end of 2026.

Before deployment, teams should still build test sets using their own Chinese accents, call codecs, overlapping multi-speaker audio, and domain-specific terminology. Artificial Analysis uses roughly eight hours of data weighted toward English-language business scenarios, while its speed measurement is based on the seven-day median for ten-minute files. The benchmark therefore cannot answer questions about time to first response for short audio, real-time streaming, upload queues, or end-to-end latency under peak concurrency. Microsoft has also not disclosed the model architecture, parameter count, weights, or complete per-language results.