Boundary conditions for cortical speech tracking as an objective speech-in-noise marker: a three-dataset MEG/EEG...

Objective electrophysiological markers are often proposed for speech-in-noise assessment, but their validation targets are not always explicit: a marker may calibrate with intelligibility, distinguish acoustic conditions, decode attention, or predict an individual speech reception threshold (SRT). We used three public MEG/EEG datasets to test these endpoints within one validation framework. In MEG-SCANS (22 normal-hearing listeners), delta-band sentence-level envelope reconstruction increased with target intelligibility (Friedman chi-square=31.27, p = 8.3e-6). However, individual curves were rarely strictly monotonic (1/22 participants), and fitted neurometric SRTs showed limited agreement with a 50% intelligibility SNR proxy (r = 0.29, p = 0.20; RMSE=2.57 dB). In the Etard/Reichenbach EEG dataset, strict leave-one-part-out decoding confirmed delta-band sensitivity to clean versus babble speech in English (p = 0.0013) and Dutch (p = 7.1e-6), with stronger target- than distractor-envelope tracking. In ds-eeg-snhl, attended speech was tracked more strongly than ignored speech (p = 6.0e-21), hearing-impaired listeners showed enhanced tracking, and the audiogram-adjusted association with SRT persisted when age was omitted. Yet EEG markers produced only small and model-dependent cross-validated predictive gains. The benchmark therefore distinguishes design-manageable constraints from construct-level non-specificity: cortical speech tracking generalizes as a group-level, condition-sensitive and attention-sensitive marker, whereas individual neurometric SRT translation remains protocol-dependent.