Abstract
People routinely name a singer they would ideally sound like. Turning that named exemplar into a measurable acoustic target, and asking whether a person’s verbal concept of a gendered voice predicts the sound they select, requires a reproducible bridge from catalog metadata to validated vocal acoustics and an evaluation protocol that cannot be satisfied by memorizing which artists appear in training. I describe both. A recording-to-representation pipeline (exemplar resolution, preview retrieval, vocal-stem isolation, eGeMAPS and singing-specific feature extraction, and artist-mean aggregation with reliability filtering) yielded acoustic profiles for 113 artists from 1,690 tracks, linked to 1,546 of 2,536 participants. In a discovery, confirmation, and hold-out design specified in a dated analysis plan, the desired direction of vocal change predicted the acoustic gender profile of the chosen singer, but its generalization to artists never seen in training did not meet the plan-specified criterion once a leakage step was corrected (retention 67.2%, 95% CI [.23, .88]), whereas participants’ specific verbal descriptions added no acoustic information beyond gender category and popularity (a bounded null). A contrastive two-tower model, retrained under a corrected protocol, also recovered no structure beyond that category. I provide the protocol in reusable form: blocking by the repeating stimulus, a retention ratio that diagnoses memorization, a chance-corrected retrieval metric, a leakage audit, and a runnable worked example.