Press play and watch Sro answer.
Every number here is its actual output.
Every recording here is real, and none was used to train Sro. The answers are Sro's readings on each clip, computed in advance and replayed in time with the audio; each one covers the seven seconds before it. "Measured" is Sro's accuracy on that kind of question over audio it never trained on. "Unmeasured" questions have no published accuracy yet.
A customer calls an AI phone agent about a blender that arrived cracked. The agent's stack sees a transcript. Sro hears the call: the moment the agent asks a question, the moment the caller cuts in over it, and the phone line it all comes through.
Each answer is a number your agent can act on directly, such as yielding the floor when two people talk at once.
Four seconds into a call about a declined card, Sro reads the caller as upset.
A quality team can ask that question of every recorded call and go straight to the ones where it holds, without anyone listening first.
Ask a speech archive who is speaking, in which language, and when the audience reacts.
Sro was never measured on crowds. It still hears the applause after "we choose to go to the moon", and the answer comes back marked unmeasured, as every new question does.
Open a live session with your questions, stream 16 kHz audio over a WebSocket, and a decision message arrives every half second. The messages below are Sro's real readings on this clip, one per second, in the same shape the live API sends.
John F. Kennedy at Rice University, 1962: public domain, John F. Kennedy Presidential Library. Support calls recorded by our team with an AI phone agent. Answers from Sro run 12, read 2 November 2026.