Personalised service and sentiment analysis for intelligent customer relationship management systems based on natural language processing

To enhance acoustic affect recognition, this study proposes the ECAPA-BGTDNN framework, which combines SE-Res2Block and Bi-GRU.It first extracts speech features via MFCC, enhances feature expressiveness through SE-Res2Block, captures temporal attributes using Bi-GRU, and achieves high-precision emotion classification via pooling and classification algorithms.Moreover, the recognised emotional states are further transformed into adaptive service strategies, enabling intelligent and real-time adjustment of customer interaction.Experimental results demonstrate that ECAPA-BGTDNN achieves 0.706 precision, 0.729 recall, and 0.717 F1-score, outperforming ECAPA-TDNN by 2.0%, 3.2%, and 2.3%, respectively.In practical deployment on a self-constructed customer voice set, the model further attains 0.837 average precision, enabling stable affect tracking during full conversations exceeding 50 seconds.