A Study on Improving the Performance of Contact Centers' Speech Recognition Using Simulated Data

초록

Contact center speech recognition enhances service efficiency. However, developing an ASR system for contact centers presents several challenges. Because of the long-form nature of contact center speech data, it is typically segmented into individual sentences for model training. However, ASR models trained on such segmented data often show degraded performance when transcribing real-world contact center speech. To address this issue, we propose two simple yet effective data augmentation methods: simulated downsizing and simulated randomness. The generated data enables ASR models to better handle original-form contact center speech. We evaluate the effectiveness of the proposed augmentation techniques on publicly available AI-HUB contact center speech datasets. Experimental results demonstrate that both methods significantly improve the recognition performance of ASR models on original-form contact center speech. © 2025 IEEE.

키워드