Sound Event Localization and Detection (SELD) integrates Sound Event Detection (SED) and Direction-of-Arrival Estimation (DOAE) to recognize and localize sound events in various applications, including urban sound sensing, wildlife monitoring, and home surveillance. Recently, advancements in machine learning, particularly deep learning techniques, have demonstrated remarkable success in improving SELD performance. However, training deep learning models for SELD is challenged by the limited availability of high-quality spatial audio data, which is essential for accurate model generalization. This paper explores the effectiveness of data augmentation techniques in overcoming this limitation. We evaluate the impact of Frequency Shift (FS), Random Cutout (RC), and Channel Swapping (CS) on SELD performance using a comprehensive set of experiments. Our findings indicate that all tested augmentation combinations except FS alone significantly improve SELD performance, reducing the SELD error by approximately 8% compared to no augmentation. The differences among effective combinations are not statistically significant, suggesting that the decision to augment is more impactful than the specific combination chosen. This work highlights the critical role of data augmentation in enhancing SELD systems and suggests future research directions, including testing these techniques with different model architectures and exploring additional augmentation methods.
Santamaria et al. (Sun,) studied this question.