Introducing Universal-3.6 Pro Realtime: Short Answers, Heard Right
Universal-3.6 Pro Realtime is live. It leads two independent streaming speech benchmarks: it sits on the Pareto frontier in Daily's Pipecat STT benchmark, and it has the lowest word error rate of any streaming model on Coval's voice-agent leaderboard.
Built for the moments voice agents get wrong:
→ Short answers like "No" and "Nah," heard right 98.5% of the time, even in loud rooms and on phone lines
→ Background voices from the TV or the next desk stay out of the transcript
→ 32 languages from one endpoint, with automatic detection and mid-sentence switching
→ Entity-aware endpointing that w