Scobleizer says Wispr Flow desktop users fix 13% fewer dictations following the launch of Canto 1, the company’s own speech model. In his November 15, 2026 post, he said the model had shipped about two months earlier. Source
Talking instead of typing loses some appeal when the computer hands you a proofreading assignment.
Separately, rohanpaul_ai wrote that dictations taking more than one second to appear were down 54% since launch. That describes fewer slow dictations, not every dictation becoming 54% faster. Source
The figures aren’t independently verified here. Neither supplied post text provides baseline rates, sample sizes for these usage figures or precise measurement windows. The 13% figure describes fewer dictations being fixed, not a 13-percentage-point jump in accuracy. Source Source
According to rohanpaul_ai, Canto 1 was trained on noisy, real-world dictations and tuned to users’ vocabulary and names. Source
The post says Wispr’s reinforcement-learning stage uses corrections from Flow users who opted in to share data. In that account, Wispr compares audio with text and examines the edits, incorporating misheard words into the training reference and ignoring ordinary sentence rewrites. Source
Changing your mind about a sentence isn’t a speech-recognition error. Neither post establishes that this training caused the reported usage gains. Source Source
Comments