Genuine and spoofed speech pairs from speakers who consented to the spoof

A detector trained on genuine speech from one corpus and spoofed speech from another learns to separate corpora, not attacks. Public anti-spoofing benchmarks repeat the problem at scale: they lean heavily on English, studio-grade recordings, and a fixed list of known attack systems, so models that top a leaderboard stumble on the first unfamiliar generator. Production fraud does not announce its generator, and it arrives over telephony.

Here the pair is the unit of collection. Every spoofed utterance is keyed to a genuine utterance from the same speaker, reading the same prompt over the same channel, so the only difference left for the detector to model is the spoof itself. That matching is what makes error analysis say something useful: when the detector fails, you know it failed on the artifact, not on a microphone.

The spoof side is scoped per project rather than fixed. TTS and voice-conversion spoofs are generated from the consented genuine recordings, covering the synthesis families your synthetic-voice detector has to catch. Replay attacks are captured physically: genuine recordings played back through real loudspeakers and handsets and re-recorded in real rooms, which is the presentation attack a microphone actually sees.

Channels get the same treatment, because a fraud call reaches the bank through a codec chain, not a studio. Genuine and spoofed material can both be passed through live telephony and VoIP paths, or injected directly at the channel, so the detector meets spoofs the way production traffic delivers them. Languages are recruited to the deployment, not inherited from an English benchmark.

Every spoofed voice in the corpus belongs to a speaker who agreed to exactly that. Participants sign use-specific cloning and biometric consent that names spoof generation for detector training and evaluation as the purpose, in plain language, and the signed chain is delivered with the data. We do not clone a voice whose owner has not signed, for any customer or any reason.

The same discipline applies to evaluation. Speakers are partitioned into held-out splits before any spoof is generated, so no test speaker's genuine or spoofed audio ever appears in training, and the reported numbers describe generalization rather than memorization. That is the difference between a benchmark score and a detector you can put in front of fraud.