This article is a continuation of research focused on the classification of EEG inner speech patterns. In a previous study, a cascade machine learning model was presented for classifying seven mental words, achieving an overall accuracy of 35.7% using a Russian-language database collected from 13 participants. In the present study, the database has been expanded to include Spanish-language EEG patterns of imagined word pronunciation. The primary focus was on modifying the data preprocessing pipeline and implementing coherence analysis between EEG channels as a key feature for classification. The study provides a detailed description of the processing steps for the Spanish-language dataset, analysis of frequency coherence levels, and the criteria based on the statistical significance of differences between baseline data and imagined word pronunciation for informative EEG channels selection. The classification performed using coherence-based features, a linear SVM classifier, and cross-validation demonstrated an average accuracy of 75.43% and 75.03% across participants for the Russian and Spanish datasets. This significantly exceeds previous results and highlights the potential of coherence as a marker for recognizing internal speech processes.