Canto: A speech model built for the real world

Wispr AI Lab introduced Canto, a new speech recognition model designed for real-world dictation conditions like noisy offices and commutes, which achieved the lowest word error rate compared to competitors from Google, OpenAI, AssemblyAI, and Deepgram on real user data. The model was trained on 10 hours of actual Wispr Flow dictations from over 2,300 speakers and tested on challenging audio conditions including background noise, music, and whispered speech.