Clipped acoustic fits did not pass a shared 1% voice pattern error test in 20 iEEG patients and 16 fMRI adults

We tested whether acoustic fits learned from non-vocal source labels predict response differences among non-speech human voices. Separate public studies supplied signed field potentials from 20 reserved iEEG patients and BOLD estimates from 16 reserved fMRI adults. Each person had a separate fit. Category and scaled spatial means were removed. Four development people per modality selected the ridge penalty before the reserved test. Mean error fell by 1.75% and 11.30% relative to centered target energy. Joint person and test-sound 95% ranges were 0.21% to 3.30% and 6.55% to 15.76%. The fixed shared rule required a gain above 1% in both modalities. It did not pass because the iEEG range extended below that margin. Resampling training sounds gave an iEEG range of -0.81% to 2.85%. Without acoustic clipping, all 20 iEEG scores were negative; fMRI changed little. An exploratory feature audit found a target electrical-source descriptor about 497 training standard deviations from its training mean. The result provides a restricted prediction benchmark, not a stable general acoustic code or a causal claim. The package preserves code, permitted inputs, source pins, all outcomes, plans, figures, and repeat instructions. Original unlicensed author files remain external.

human auditory cortex · iEEG · fMRI · acoustic encoding · held-out sound class · prediction error · extrapolation