Sound whose truth is known: a reference material for acoustic voice measurement

The data descriptor for reference material v0.7: what is in it, how it was built, what it is valid for, and what it cannot do. 930 synthetic specimens across five task classes with exactly known truth, generated deterministically from seeds, each with a SHA-256 over its decoded samples. The technical validation reported here was run on 2 September 2026: all 930 digests verify, and a standard implementation (Praat through Parselmouth) was measured against the stated truth on the 186 sustained-vowel specimens, on the clean condition and across all 31 recording conditions. On the clean specimens the measured value tracks the truth closely for fundamental frequency (r = +0.986, slope 0.979) and for shimmer (r = +0.970, slope 0.826), less well for jitter (slope 0.656), and not at all for harmonics-to-noise ratio (r = -0.140), which is the definitional problem this project has documented separately rather than an estimator failure. THE FIVE TESTS any new measure should report, all computable from this material and none requiring new data: response slope against the truth; drift across conditions while the truth is fixed; robustness by degradation family; refusal behaviour, including how often the measure returns nothing; and confounding, since the material can vary one property while holding the others rigidly still, which no real recording permits. WHAT IT CANNOT DO, published because an honest reference material states its own limits: it cannot imitate the period structure of a real larynx, so no precision ceiling for jitter can be derived from it; it has no physiological amplitude reference; its phonation breaks are filled with attenuated aspiration where they should contain real ambient noise; and its coverage of female and child voices is thin. A reference material should not try to be realistic. Its value is control: you can set one property and hold the others still. A material that is transparently artificial but exactly known is more useful than one that is convincingly lifelike and approximately known, and considerably safer.