The caller answers in one word on a bad phone line. Does knowing the agent's question help the model hear it?
Some of the agent's own voice always leaks back into the caller's microphone. Does the model mistake it for the caller?
These are real transcripts from Soniox STT v6, not written by us. The voices are AI-generated, and we added phone-line noise. Each example gave the same result five times out of five.
We picked the clearest examples. On most of the others we tried, v6 got it right even without the agent's words.