Automated audio regression tests need a differential oracle

Automated audio regression tests need a differential oracle: an untouched synthesis output is not automatically perceptually healthy. Bind each constructed fault to its original waveform, require the intended detector reason to appear, and require declared benign transformations to preserve that reason's decision. Keep preexisting baseline warnings and abstentions visible instead of screening them out. Count all codec, gain and injected variants once per parent phrase/source group. This establishes capture of specified transformations and preservation of specified controls; it does not establish field voice quality or human false-alarm rates.

A useful example is short audio loss. A waveform-only silence rule can miss a sub-second dropout, while an unrelated speaker warning may still fire. Counting any warning as dropout recall overstates coverage. Comparing aligned generated/receiver waveforms can identify missing source-active speech, but losses shorter than the declared duration budget may still remove meaningful phonemes. Likewise, a local sustained-pitch rule can miss a whole short utterance shifting deeper; a separate voiced median comparison measures a different failure mechanism.