ECHO: A Matched-Contrast Benchmark for Context-Sensitive Turn-Taking in Full-Duplex Dialogue
The benchmark tests whether full-duplex dialogue systems can tell the same overlap apart by context.
ECHO pairs Chinese dialogue examples with identical overlap transcripts but different preceding turns, so one case calls for yielding and the other for keeping the floor. It also adds off-talk cases to catch systems that yield when they should not. The authors report that most tested systems lean strongly toward `Yield`, doing better on interruptions than on backchannels. They argue that interruption-only testing can make turn-taking reliability look better than it is. ArXiv · AI/CL/LG's note
ECHO pairs Chinese dialogue examples with identical overlap transcripts but different preceding turns, so one case calls for yielding and the other for keeping the floor. It also adds off-talk cases to catch systems that yield when they should not. The authors report that most tested systems lean strongly toward `Yield`, doing better on interruptions than on backchannels. They argue that interruption-only testing can make turn-taking reliability look better than it is. ArXiv · AI/CL/LG's note
score 5