Details
- Microsoft AI says MAI-Transcribe-2 ranked first for transcription accuracy in David AI’s DAI-ASR-I18N benchmark.
- The model recorded the lowest error rate in 18 of the benchmark’s 21 languages.
- David AI evaluated automatic speech recognition on 147 hours of unscripted conversations across 21 languages.
- MAI-Transcribe-2 stayed below a 10% error rate in 16 languages, according to the benchmark results.
- The model is available in Microsoft Foundry, Microsoft’s platform for deploying and using AI models.
- MAI-Transcribe-2 supports 60 languages and includes speaker separation, word-level timestamps, automatic language identification, and configurable transcript styles.
- Microsoft positions the model for captioning, meetings, clinical notes, call-center records, accessibility, content creation, and voice agents.
Impact
The result strengthens Microsoft’s position in multilingual speech recognition and gives Foundry customers a measurable reason to test its model against established commercial alternatives. Its broad language coverage and benchmark performance could support adoption in global customer-service, accessibility, and media workflows, although real-world cost, latency, privacy requirements, and performance on domain-specific audio will remain decisive. The benchmark is also focused on conversational speech, so the result does not establish an across-the-board lead in every transcription use case.