# Microsoft Launches AI Speech-to-Text Model MAI-Transcribe-2

- Category: Flash (快讯)
- Published: 2026-09-04T11:54:00.000Z
- Source date: 2026-09-04
- HTML: https://www.88z.ai/en/news/flash-20260904-6195913078d6fc49
- Markdown: https://www.88z.ai/en/news/flash-20260904-6195913078d6fc49.md
- Other language: https://www.88z.ai/news/flash-20260904-6195913078d6fc49.md

Microsoft has released its latest AI speech-to-text model, MAI-Transcribe-2. In FLEURS tests across 60 languages, it achieved an average word error rate of just 5.2%, significantly outperforming competitors like Gemini, GPT-Transcribe, and Whisper in both accuracy and speed. The model supports intelligent speaker differentiation, word-level timestamping, flexible switching between verbatim and clean styles, and natural code-switching in conversations.

## At a glance

- Achieved a 5.2% average word error rate in 60-language tests, outperforming competitors
- Supports speaker differentiation, timestamp annotation, and mixed-language conversation processing

## The update

Microsoft has officially launched MAI-Transcribe-2, its most powerful AI speech-to-text model to date. In FLEURS benchmark tests covering 60 languages, the model demonstrated exceptional performance with an average word error rate of only 5.2%. Data indicates that it significantly leads mainstream competitors such as Gemini, GPT-Transcribe, and Whisper in both accuracy and processing speed. In terms of functionality, MAI-Transcribe-2 features intelligent speaker differentiation and provides precise word-level timestamp annotations. Users can flexibly switch between verbatim and clean output styles depending on the specific application scenario. Furthermore, the model supports the processing of naturally code-switched conversations, allowing it to better adapt to complex multilingual communication environments.

## Sources

- [IT之家](https://www.ithome.com/0/998/241.htm) (2026-09-04)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
