Measurement data quality in room acoustics: A systematic audit framework for field campaign errors and their quantified impact on machine learning prediction performance

Acoustic impulse response measurement campaigns conducted in performance venues under ISO 3382-1 are susceptible to a class of systematic field errors, including receiver position labelling swaps, spatial co-ordinate entry mistakes, signal-to-noise ratio failures at low frequencies, and level calibration interruptions, that have received little systematic documentation and whose downstream impact on quantitative analyses has not previously been characterised. This study presents a systematic data quality audit framework demonstrated on a 171-seat conference hall (Başöğretmen Atatürk Conference Hall, V = 1511 m ³ ); the methodology is applicable in principle to any ISO 3382-1 campaign pending multi-venue validation. Six independent error categories were identified and corrected through cross-checking between recorded source-receiver distances, architectural co-ordinate databases, Barron’s revised room acoustics theory, and physical parameter range constraints. Five of these errors were genuine field issues discovered through the audit; one (the G calibration interruption) was deliberately introduced as a blinded controlled-validation case to verify the detection mechanism against known ground truth. The impact of each error category on downstream machine learning prediction accuracy was quantified using a gradient boosting ensemble framework with cross-validated R ² and held-out validation R ² as performance metrics. Correcting these errors improved downstream cross-validated prediction performance, most clearly for the spatially-driven parameters: cross-validated R ² for T ₃₀ and ST rose by amounts exceeding the cross-validation uncertainty, and the calibration correction improved G prediction. Given the single-venue sample (54 training rows, 3-fold cross-validation, with fold standard deviations of 0.2–0.5), these effects are reported as directional evidence that undetected field errors propagate into spatial prediction models, rather than as precise effect sizes. The audit framework is formalised as six detection algorithms and an on-site verification protocol, providing a systematic illustration of how each measurement error category propagates through a machine learning acoustic prediction pipeline. All quantitative results derive from a single 171-seat conference hall; the reported Δ R ² values are venue-specific effect sizes that carry substantial cross-validation uncertainty and await cross-venue replication.