Emotion recognition in conversation (ERC) underpins agent‑assist prompts, escalation routing, and post‑call analytics in contact‑center‑as‑a‑service platforms, where cost and latency are as critical as accuracy. We compare three deployment options at the system level: (1) a low‑cost stacked ensemble that combines sentence embeddings, windowed context, and stacking of RandomForest, XGBoost, and Logistic‑Regression; (2) off‑the‑shelf LLM prompting using GPT‑4o‑mini in zero‑shot, few‑shot, and chain‑of‑thought modes; (3) a confidence‑gated hybrid that escalates only the ensemble’s least‑confident predictions to the LLM, mirroring IVA‑to‑human escalation policies in production contact centers. On the IEMOCAP benchmark, the ensemble achieves a weighted F1 of 0.595, significantly outperforming all LLM configurations (0.460‑0.536, p<0.05). In terms of latency and cost, the ensemble runs in ~30 ms per utterance at roughly one‑tenth the cost of the LLM; the hybrid adds less than 5% cost by invoking the LLM on only about 15% of low‑confidence samples while preserving near‑ensemble accuracy. These findings suggest that a lightweight ensemble with confidence‑gated LLM escalation offers a practical balance of accuracy, latency, and expense for production‑grade conversational AI.
Review: The work delivers actionable deployment guidance and demonstrates a cost‑effective path for real‑world contact‑center deployments.