MAI-Transcribe-1.5

MAI-Transcribe-1.5 is Microsoft AI’s updated speech-to-text model, adding entity biasing to the production-grade transcription accuracy of MAI-Transcribe-1. Developers prime the model with a list of domain-specific entities — meeting attendees, product names, medical vocabulary, scientific terms — and the model biases its predictions toward those entries when the acoustic signal is ambiguous. The system retains its #1 spot on the FLEURS benchmark with Word Error Rate improving from 3.9% to 3.7%, and on Artificial Analysis it reaches 2.5% overall WER at rank 3, handling cross-talk, background noise, and long-form meetings in production conditions.

The 1.5 generation also doubles down on the model’s speed and cost story: throughput is up roughly 3× and latency improved by 5.7×, letting the model transcribe an hour of audio in under 10 seconds — more than 5× faster than Gemini 3.1 Flash, ScribeV2, and gpt-4o-transcribe. The combination of best-in-class accuracy on multilingual benchmarks, content-aware biasing, and frontier-leading throughput positions MAI-Transcribe-1.5 as the production transcription layer for meetings, voice agents, and enterprise content workflows — where speed and domain accuracy matter as much as raw WER.