The study of speech in natural environments poses challenges for traditional electroencephalogram (EEG) analysis approaches. In recent years, machine learning models—particularly regularized linear encoding models—have enabled a transition toward experimental designs that incorporate dynamic and naturalistic stimuli, such as speech during dialogue. This work aims to understand how different speech attributes are encoded in the brain within the context of unscripted natural dialogue. To this end, we extract low-level attributes (envelope, pitch, spectrogram, among others), high-level attributes (phonemes, phonological features, among others), and attributes derived from representations obtained with deep neural networks (Wav2Vec2.0, Whisper). The results show that the inclusion of high-level attributes significantly improves the prediction of brain signals across all frequency bands. In particular, predictions based on phonemes and phonological features suggest that neural sensitivity is consistent with the hypothesis of a hierarchical language processing system.