The use of brain-computer interfaces to reconstruct intended speech from neural activity offers hope for individuals with severe speech production impairments. While decoding overt speech (speech that is spoken out loud) has progressed, decoding imagined speech has been more challenging.To address this, electrocorticography data was collected from 13 patients who performed both overt and imagined speech tasks. Consistent and specific neural features were extracted from the data, and their effectiveness in differentiating speech items in articulatory, phonetic, and vocalic spaces was evaluated. While high-frequency activity was most effective for decoding overt speech, low- and higher-frequency power and local cross-frequency contributed to the decoding of imagined speech, particularly in phonetic and vocalic spaces.