Abstract
The German Fach system divides operatic voices into roughly 25 categories within the familiar voice classes, distinguished by tessitura, vocal weight, timbre and agility. Whether these categories are acoustically recoverable has rarely been tested on singers a model has not seen. I analysed 8,681 public-domain 78-rpm recordings by 508 singers in 27 Fächer, labelling each singer with their documented career Fach. After a design audit, 153 singers were locked aside, 12 tests were specified in advance, and the frozen pipeline was evaluated once on the held-out singers. Seven-way voice class was recovered at a balanced accuracy of .54 (chance .15), and female-range versus male-range class at .97 without explicit pitch measures. Within a voice class, the Fach label did not generalise: tenor sub-Fach classification fell from .54 on discovery singers to .38 on held-out singers (chance .34), and no single-aspect model replicated. Treated as a continuous position on the taxonomy's tessitura ordering, however, soprano Fach was predicted on held-out singers (ρ = .44). Female-range voices in electrical-era recordings showed lower singer's-formant prominence and a higher Hammarberg index than male-range voices, differences that a synthetic audit showed are not measurement artefacts of high pitch. The Fach of the role being sung was more recoverable than the singer's label. Coarse voice class is acoustically robust; within a class, the evidence favours an ordering over the discrete categories the labels draw, and points to repertoire as part of what the labels encode.