Spatial sampling requirements for data-driven prediction of room acoustic parameters: A within-venue sub-sampling experiment

How many measurement positions does data-driven spatial characterisation of a performance venue require? A controlled within-venue experiment answers this question, distinguishing throughout between two often-conflated tasks: predicting acoustic parameters at positions never measured and transferring measured positions across environmental conditions. In a 1077 m 3 stepped conference hall (264 seats), 75 receiver positions were measured under varying temperature and occupancy, including conditions with seated audiences. Models were trained on random subsets of 15–60 positions and evaluated on the held-out remainder, with inverse-distance interpolation on identical splits; 11 ISO 3382-1 parameters were predicted from source–receiver geometry and environmental descriptors. The density–accuracy relationship at unmeasured positions is strongly parameter-class dependent. Sound strength is predictable from geometry at all densities and shows the most consistent learning advantage over interpolation; reverberation time is the only parameter whose accuracy grows visibly with density; clarity and temporal parameters remain modest at every density and are matched by interpolation beyond sparse grids. Perceptual and statistical rankings are nearly inverted: centre time, statistically modest, is the only parameter predicted within its just-noticeable difference, while sound strength, statistically strongest, carries the largest perceptual error. At measured positions, condition interpolation attains the high accuracies reported in the literature. Cross-comparison with a 180-seat shoebox concert hall reveals typology-specific environmental sensitivity: at matched positions, occupancy and temperature measurably affect the conference hall but leave the purpose-designed concert hall essentially unchanged. Sampling density should be specified per parameter class and per question; these results provide a quantitative basis for doing so.