Freeze the files before judging them
A reproducible comparison starts with one source excerpt, named model builds, fixed settings, and downloadable or inspectable outputs. If any clip is trimmed, normalised, compressed, or made louder, the record should say so. Otherwise the presentation can influence the result more than the separator.
A metric answers one narrow question
Signal-to-distortion ratio compares an estimate with a known reference signal and expresses the relationship in decibels. Scale-invariant SDR removes constant gain from that comparison. These measurements are useful, but one number cannot describe every musical artefact or every listening purpose.
Listening needs a protocol too
Candidate clips should be anonymised, level matched, and presented in a controlled order with a precise prompt. Listeners may judge interference, artefacts, naturalness, or usefulness for practice. The published result should include the question, participant count, aggregation method, and known limitations.