For fixed sound panels, class labels lowered mean error beyond acoustic class scores in iEEG and fMRI

We tested whether acoustic class scores learned from training labels can replace known target labels when predicting auditory responses. Separate public studies supplied signed LFP means and BOLD fMRI. Development used four people per modality and excluded reserved target responses. A local lock preceded testing in 18 patients and 16 MRI adults. Known labels reduced mean error beyond acoustic class scores by 2.10 and 9.04 reference-error points. Their 95% person Student limits were [0.53,3.68] and [8.42,9.67]. A shared rule also required gains over acoustics and a constant. It passed correction over two claims, Holm p=0.04316. The training-only acoustic-supervision claim failed, p=0.9062. An added eight-adult MRI study gave a label gain of 8.91 points. Crossed person and sound resampling gave an iEEG range of [-0.56,5.56], so transfer to arbitrary new sound samples was not confirmed. Six music clips supplied the positive iEEG class average. Removing vocal sounds left an electrical comparison uncertain. Pre-event subtraction left overall electrical skill uncertain. An exact omitted-acoustic-variable model gave a label advantage without an abstract category response. All 110,682 electrical means matched source chunks. Two arithmetic forms agreed on 42 checked people. One earlier MRI stored table differed slightly; its cause remains unresolved and both records remain. The correction changed no test decision. Clinical sampling, source selection, different libraries and measurement filters limit scope. Code, licensed compact inputs, pins, all outcomes, figures and repeat instructions are supplied. No common neural mechanism or cause is established.

auditory cortex · iEEG · fMRI · sound categories · acoustic features · kernel regression · held-out prediction · person and sound uncertainty