Express: The Sound of Words: Verbal and Vocal Cues in Customer Satisfaction

Technology-mediated service interactions, such as phone calls, generate rich behavioral signals that firms currently struggle to monitor in real time. Post-interaction surveys provide delayed and unrepresentative feedback, leaving satisfaction invisible as calls unfold. Satisfaction judgments depend on how evaluative meaning is communicated across two channels, verbal content and vocal delivery, whose relative contribution varies by outcome type. The authors analyze 1,144 real-world service calls, each coded as dissatisfied, neutral, or satisfied, using text-only, audio-only, and multimodal deep learning models. Dissatisfied calls are identified from transcripts alone as accurately as by multimodal models, since customers articulate problems and remedy expectations explicitly. Neutral and satisfied calls depend more strongly on vocal interpretation, and cross-attention models outperform text-only alternatives precisely where verbal content is absent or insufficiently distinctive to separate outcomes. The findings identify a boundary condition on interactivity, whereby vocal cues add predictive value when evaluative meaning is absent or insufficiently distinctive between outcomes, not when it is explicit and clearly distinguishes outcomes, as is the case for dissatisfied calls.