Disentangle bona fide-domain and synthetic-output effects in detector adaptation

Determine whether the difference in legacy-benchmark performance between LibriSpeech- and VCTK-adapted XLS-R SLS detectors is caused by the bona fide speech domain, the paired synthetic outputs, or both.

Background

The detector-adaptation experiments use nearly equal-sized LibriSpeech and VCTK adaptation sets with the same number of updates, but the resulting detectors differ in their performance on legacy benchmarks. VCTK lies within the bona fide domain used to train the original detector, whereas LibriSpeech does not. Because changing the adaptation corpus simultaneously changes both the bona fide utterances and the synthetic outputs paired with them, the experiment does not isolate which factor causes the performance difference.

References

However, switching the corpus changes both the bona fide utterances and the synthetic outputs paired with them, so this experiment cannot distinguish the effect of the bona fide domain from that of the synthetic outputs.

— Tracing and Relearning Detection Evidence in Text-to-Speech Systems  (2609.30983 - Shin et al., 25 Sep 2026) in Section 4, subsection “Adaptation learns tuned outputs”