Intelligent Call Recording Quality Inspection Using Dual‑Mode Detection

Background: Many enterprises operate large call‑center operations where communication quality directly impacts revenue and customer satisfaction. Existing keyword‑and‑regex based quality inspection tools suffer from severe false‑negative and false‑positive rates, failing to meet the growing demand for precise and efficient monitoring.

Goal: To address these challenges, the paper presents a dual‑mode detection framework that automatically monitors key scripts, red‑line words, and semantically similar utterances in customer‑service recordings, ensuring high precision, rapid response to new tasks, and robust detection of varied expressions of standard scripts.

Preparation – Similar‑Semantic Library Construction: Using business‑provided standard scripts (e.g., “How are the other subjects for the child?”), the HANLP library performs word segmentation, then a synonym tool generates synonyms for each token. All synonym combinations are enumerated to create a comprehensive similar‑semantic sentence library.

Preparation – Sentence Segmentation Model: A BERT‑based sequence labeling model trained with focal loss segments ASR output into accurate sentences, mitigating the common issue of merged sentences in raw transcription.

Preparation – Semantic Classification Model: A RoBERTa multi‑class model is trained on all business‑defined intent labels (e.g., curriculum inquiry, satisfaction follow‑up, discount offers). The model achieves macro‑averaged precision 0.93, recall 0.87, and F1 0.89, offering higher generalization than rule‑only approaches.

Detection Process: 1) ASR results are first segmented into fine‑grained sentences. 2) Rule‑based detection matches segmented sentences against the similar‑semantic library; matched candidates undergo dependency‑graph analysis, and a detection is confirmed when semantic keywords share a dependency path. 3) Model‑based detection feeds the sentences into the semantic classification model for broader intent recognition. 4) Rule and model results are fused; only when both miss a sentence is it considered undetected.

Result Presentation: Detected outcomes are stored in a database and visualized on a dashboard, providing statistics for quality inspectors and business teachers.

Flowchart: The paper includes a process flow diagram illustrating the end‑to‑end pipeline from ASR input to final result fusion.

Conclusion: The proposed dual‑mode detection solution satisfies the requirements of high detection precision, rapid adaptation to new tasks, and strong generalization for customer‑service quality inspection.

References: 1) "Intelligent Quality Inspection Whitepaper" – Loop Intelligence; 2) "RoBERTa: A Robustly Optimized BERT Pretraining Approach".