AssemblyAI says Universal 3.6 Pro Realtime is live

AssemblyAI says Universal 3.6 Pro Realtime is live. The company says two independent benchmarks ran every streaming speech model on the same audio and the same harness, and that Universal 3.6 Pro Realtime led both.

On trydaily’s Pipecat open speech-to-text benchmark, AssemblyAI says the model sits on the Pareto frontier. On covaldev’s live leaderboard, the company says it has the lowest word error rate of any streaming model on real voice-agent audio.

AssemblyAI says short replies such as “No,” “Nah,” and “Nuh-uh” are heard correctly 98.5% of the time in loud rooms and on phone lines, not only on quiet calls, and that background voices from a TV, a coworker, or a nearby desk stay out of the transcript. It also says the model covers 32 languages from one endpoint with automatic detection, including callers who switch languages mid-sentence, and uses entity-aware endpointing that ends a turn when the caller is finished while holding it open during a phone number. AssemblyAI says it was trained on tens of thousands of hours of real voice-agent and telephony conversations.