Blind Pass v2

26 two-channel conversations (left = speaker A, right = speaker B), all seven defect types in play. Each clip has at most ONE planted defect — or none.

Backchannels ("yeah", "mm-hmm") and natural overlap are NOT defects. Pick a verdict per clip, add the defect onset if you heard one, then copy the export block back. Answers autosave in this browser.

Defect guide — what each verdict means

clean: Nothing planted. A natural chat — backchannels, quick tempo, even brief overlap are all normal.
response_delay: Abnormal dead air before a reply — the conversation stalls for seconds where a quick handoff should happen, then continues normally.
pause_barge_in: A speaker pauses mid-sentence (a thinking pause, not a handoff). The other speaker misreads it and lunges in to grab the floor — but the first speaker just keeps going, talking over or past them. Listen for the three beats: mid-thought pause → floor-grab → the grab gets steamrolled.
generic_filler: A reply with zero content that dodges what was just said or asked — stacked vague pleasantries ("wow, yeah, that's really something") where a real answer should be. The tell: a direct question gets left hanging.
off_topic_reply: A reply about something unrelated — grammatically fine, but it ignores the thread of the conversation and talks about a different subject.
voice_swap: One speaker's voice becomes a different person's voice partway through, while the conversation content continues as if nothing changed.
audio_artifact: Non-speech corruption on one channel — a run of sharp clicks, or a burst of static.
unsure: Something felt off but you can't place it — say what you heard in the notes.

Never flag: backchannels ("yeah", "mm-hmm" under the other speaker), natural brief overlap at turn edges, or TTS quirks (odd pacing, pronunciation, robotic prosody) — the engines differ and that's not what's being tested. Each clip has at most ONE planted defect.

0/26 answered

blind_0132s