This paper proposes a two-branch hierarchical model for the complex assessment of the foreign-language oral speech of convicted persons based on segmental and suprasegmental indicators. The segmental branch combines the Goodness of Pronunciation (GOP) obtained from forced phonetic alignment, phone accuracy, pronunciation completeness, duration and boundary stability; the suprasegmental branch computes the fundamental tone contour, word stress, energy, rhythm, pause, articulation rate and fluency. Each branch produces a separate subscore and a reliability value. A meta-level dynamic gate adjusts the channels' contributions according to audio quality, task difficulty, alignment reliability and prosodic stability; the final result consists of a 0-100 complex score, the two subscores, and an explainable error map. The model was demonstrated on 720 synthetic speech samples belonging to 120 virtual profiles, in a cross-validation scheme with six outer and five inner blocks. The proposed model achieved Pearson r=0.963, Spearman ρ=0.963, MAE=3.27 and RMSE=4.11 points. AC≥0.55filterreduced the MAE to 3.19 points at 84.6% coverage. These indicators do not pertain to real respondents; they are a reproducible synthetic verification of the model architecture and the assessment protocol.