Speech comprehension involves an active neural process mapping sounds into meaning, with lexical predictions supporting comprehension based on the preceding context. Methods like the Temporal Response Function (TRF) have proven effective in probing the neural processing of sounds properties such as the acoustic envelope based on EEG/MEG. However, examining how the human brain processes words during continuous speech listening remains challenging.This study presents an assessment of encoding TRF methodologies for the measurement of lexical processing, quantifying and explaining their limitations. Next, we introduce two novel metrics that substantially improve the assessment of lexical processing. The first metric revisits how to calculate the total EEG variance explained by lexical information, magnifying the effect by excluding temporal intervals that could not be affected by lexical surprise. The second metric relies on the weights of multivariate TRFs by taking into account the collinearity between stimulus features. All analyses were carried out on simulated EEG signals and then validated on a publicly available EEG dataset.Numerical results indicate that the proposed metrics are substantially more sensitive to lexical surprisal than the vanilla TRF evaluation, leading to effect sizes ranging from 80% to 100% increase compared to the typical TRF evaluation. Furthermore, this enhancement in effect size comes with no additional complexity or computational demands. We conclude that these new metrics will enable more thorough investigations into the underlying mechanisms of lexical predictions in natural speech comprehension.