Microsoft's MAI-Transcribe-2 tops David AI's 21-language speech benchmark

Microsoft AI's MAI-Transcribe-2 posted the lowest transcription error rate in 18 of 21 languages in David AI's public DAI-ASR-I18N evaluation. The benchmark used 147 hours of unscripted conversation, and the model stayed below 10% error in 16 languages. ElevenLabs Scribe v2 led in the remaining three languages, including English.

• MAI-Transcribe-2 leads in 18 of 21 languages by WER and CER
• Error rate stayed below 10% in 16 languages
• Benchmark covers 147 hours of conversational speech and diarization
• ElevenLabs Scribe v2 led in 3 languages, including English