Microsoft AI Introduces MAI-Transcribe-1.5: 2.4% WER on Artificial Analysis, Best-in-Class FLEURS Accuracy, and Up to 5x Faster Long-Audio Transcription

Editor's take

Microsoft AI has unveiled MAI-Transcribe-1.5, an updated speech-to-text model demonstrating a 2.4% Word Error Rate (WER) on the Artificial Analysis benchmark and improved accuracy on the FLEURS dataset. This iteration expands language support to 43 and introduces keyword biasing, a critical feature for improving transcription accuracy in specialized domains like legal or medical fields.

The significance lies in MAI-Transcribe-1.5's continued push towards parity with, and in some areas potential surpassing of, leading commercial offerings, including those from OpenAI and Google. The emphasis on speed (up to 5x faster for long audio) alongside accuracy and multilingualism positions it as a potent internal tool for Microsoft's vast product ecosystem and a potential competitive offering in the enterprise AI services market.

Future developments to monitor include MAI-Transcribe-1.5's real-world performance across diverse accents and noisy environments, beyond benchmark datasets. The extent to which its keyword biasing can be effectively deployed by end-users without extensive technical expertise will be crucial for its broad adoption and impact on transcription service providers.