When a human-labeled call contains AI speech
Listening adjudication uncovered screeners, voicemail, and recorded prompts inside the negative class. Why field false-positive claims fail without ears on the audio.
IN REVIEW01RESEARCH
From the lab.
Measurements, null results, and what synthetic-voice detection actually looks like on the phone network. Written by the people who ran the evaluations.
STATUS
The field-evaluation manuscript and supporting research notes are in review. No article ships before its claims clear the same evidence discipline as the rest of this site.
What replication changed — and what it did not
A fresh, zero-overlap production cohort reproduced the fielded result while the raw human-label flag rate moved. The case for frozen operating points and adjudicated metrics.
IN REVIEWThe detector that earns its slot without detecting
Removing the third family destabilized the fused score geometry. Its value is distributional anchoring, not standalone discrimination.
IN REVIEWOpposite failure modes: why our ensemble is load-bearing, not marketing
Telephony conditioning breaks the two strongest detector architectures in opposite directions. The measurements that made a third family non-optional.
IN REVIEWTwo null results that found the real domain gap
Neither synthetic codec conditioning nor real PSTN transmission reproduced our false positives. Spontaneous speech did. Domain gaps live in speech style, not codec math.
IN REVIEWYour lab-perfect detector is wrong on the phone network
Why checkpoints selected on benchmark metrics produced telephony false positives, and what selecting on deployment audio looks like.
IN REVIEW