Metric optimization for Sound Event Localization and Detection
Soo‐Young Park, Youngho Jeong, Taejin Lee · 2020
Due to DNN-based approach, sound event localization and detection is now possible with a simple microphone array. However, this DNN-based sound event localization and detection has problem with dataset imbalance, pooling size decision, and periodicity of angles. In this paper, we propose three methods: soft f-loss with temporal masking, periodic loss, and PoolNet-based architecture to handle three issues. We evaluate our approach on TAU-NIGENS Spatial Sound Events 2020 dataset. The experimental results show our approach shows performance improvement in sound event localization and detection metrics compared to the baseline system.