Sentiment expressed in naturalistic conversation is carried jointly by what a speaker says, how their voice sounds, and how their face moves, yet most automated sentiment systems still rely on a single channel. We recorded 186 adults completing semi-structured autobiographical narrative interviews, yielding 5,412 utterance-level segments independently labelled by three trained annotators for categorical sentiment (positive, neutral, negative; mean pairwise Cohen's kappa=0.72) and continuous valence-arousal ratings. We extracted contextual text embeddings, prosodic and spectral audio features, and facial action-unit dynamics for every segment, and compared unimodal classifiers against early fusion, late fusion, and a proposed cross-modal attention transformer (CMAT) that lets each modality attend to the other two before classification. Unimodal performance was modest (text macro-F1=0.61, audio=0.54, facial=0.52), while fusion improved classification substantially, with CMAT reaching the highest macro-F1 of 0.75 and 79.8% accuracy, significantly outperforming late fusion (McNemar p=0.012) and text-only classification (p<0.001). Leave-one-modality-out ablation of the fusion model showed text contributed most to performance (removal dropped macro-F1 to 0.64), but audio and facial removal also produced reliable, smaller decrements (macro-F1 0.71 and 0.72, respectively), indicating genuine complementary information rather than redundancy. CMAT-predicted continuous valence correlated strongly with human ratings (r=0.74), exceeding any unimodal model, and learned cross-modal attention weighted the facial channel more heavily during high-arousal negative utterances than during neutral ones (r=0.36 between facial attention weight and rated arousal). These results indicate that text, audio, and facial cues carry partially independent sentiment-relevant information, and that transformer-based cross-modal attention recovers more of this information than fixed early or late fusion schemes, with implications for affect-sensitive human-computer interaction systems.