The joint modeling of mispronunciation detection and diagnosis (MDD) and automatic pronunciation assessment (APA) in computer-assisted pronunciation training (CAPT) systems has been proven effective. While existing approaches (e.g., HMamba) employ self-supervised learning (SSL) to extract rich speech representations incorporating prosodic features like phoneme duration and silent segments, they still face two key challenges: On one hand, SSL representations lack targeted modeling of crucial suprasegmental features such as stress; on the other hand, directly joint-training segmental features (local phoneme accuracy) and suprasegmental features (global prosodic patterns) leads to performance conflicts due to their differing granularities of focus. To address these issues, this paper proposes a stress-enhanced framework. First, we explicitly model word-level and sentence-level stress features based on vowel formants (F1/F2) and spectral balance. Subsequently, we employ a proficiency-aware attention matching mechanism that adaptively adjusts fusion weights between segmental and suprasegmental features according to learners’ second-language proficiency.