Backwards compatibility to classical experiments grounds beta responses to naturalistic speech in temporal acoustic forecasting

Current neuroscience is shifting from simple controlled paradigms towards rich and ecologically valid naturalistic stimuli. Correspondingly, insights from historic "impoverished" artificial paradigms are considered to be seriously challenged by generalisation to modern naturalistic contexts. Here, we argue that the reverse, backwards compatibility, is an under-appreciated and easily accessible benchmark and show that it disambiguates model comparison beyond naturalistic stimuli. We analyse magnetoencephalography (MEG) beta power responses to an audiobook stimulus using canonical correlation analysis (CCA) and replicate reported links between beta power and syntactic parsing. However, a simple stimulus-computable acoustic model predicts the same variance, suggesting a domain-general rather than linguistic function of beta responses. We therefore test the backwards compatibility of speech-trained models to classic rhythmic tones. While generalisation initially fails, reducing hidden and unnecessary degrees of freedom of the models' phase responses allows successful generalisation. Crucially, this greatly improves model adjudication: Several models that perform indistinguishably on speech differ in how well they predict responses to simple tones. In this comparison, a simple forecasting deep neural network (DNN) outperforms acoustics by internalising a "slow-decay" prior as a structural mirror of sluggish speech dynamics. This grounds beta responses in canonical temporal forecasting, bridging modern naturalism to established auditory psychophysics.