Wispr Flow

Recently shipped Canto, a proprietary real-time speech model now running all dictation automatically.

Canto beats the third-party stack on messy audio. Measured on 9.8 hours of real-world English recordings from 2,300+ speakers, it posted the lowest word error rate of every model tested, Google, OpenAI, AssemblyAI, and Deepgram included. It also led on stress-test audio: background conversation, music, traffic, wind, whispered speech. A 10x-scale successor is in training, targeting harder conditions and broader language coverage.

Notetaker reached Windows and pulls in outside AI tools. Windows joins Mac, with meeting detection, personal-dictionary term handling, in-call catch-up, cross-meeting search with citations, and draft generation. Gemini joined Claude and ChatGPT in one-click setup, and any MCP-compatible tool can now read transcripts — connection grants transcript access only.

Meeting capture got sturdier.

• Speaker detection in Teams gallery view and Google Meet
• Stuck Windows meeting uploads now finish on their own, and recordings recover from save or upload errors
• @-mention people in meeting notes, with hover contact cards
• A warning when the mic picks up no sound, plus a “Speaking a different language?” prompt to add languages on Mac and Windows

Privacy controls tightened. Model training stays off for Growth and Enterprise members. Turning off dictation Cloud Storage no longer stops Scratchpad from syncing across devices.