Xiaomi's open model picks one voice out of three talking at once.
Xiaomi released the free CocktailASR-1 speech model on Friday. You give it a short sample of the voice you want, and it ignores everyone else. On a three-voice test it got 12.29% of words wrong, against 4.11% with two voices.
Why it mattersMeeting notes and captions fall apart the moment two people talk over each other. Open weights mean this can run on your own machine, not a company's server.
▶ Related video: Solving the Cocktail Party Problem with Deep Learning (In-depth version)
Read in the app — free, in 9 languages
Related stories
DeepSeek is testing its first voice mode on a slice of users.2026-09-14ElevenLabs' new music model lets free users sell what it writes.
2026-09-13A Chinese retailer's free model draws the world as you walk into it.
2026-09-13Kimi's cheap coding model now opens a million-token window to all.
2026-09-12A Chinese map app built a 3D world you can fly through.
2026-09-11