A Lightweight Two-stage Sound Event Localization and Detection Method Based on Convolutional Networks

For sound event localization and detection (SELD), neural networks are often used to identify sound events and estimate their location. However, to achieve optimal prediction results, the required neural networks are becoming larger and more complex, leading to increasing hardware demands that limit their application in mobile terminals. In this paper, we propose an efficient SELD task estimation method that uses gated depthwise separable convolution (GDSC) for feature extraction. This method is applied in a two-stage SELD structure to transfer the feature layers learnt by sound event detection (SED) to direction of arrival (DOA) estimation. In addition, different strategies are used to deal with time series data in both the SED and DOA branches. Gated recurrent unit (GRU) is used in the SED branch, while temporal convolutional network (TCN) is included in the DOA branch to enhance SELD task’s accuracy. We evaluate our proposed approach on the DCASE 2019 Task 3 multichannel microphone array (MIC) dataset. According to the experimental results, the proposed GDT-SELDnet method achieves an error rate of 0.154, a F-score of 91.1%, a DOA error of 10.614°, and a frame recall of 86.5% for the SELD task with a SELD error of 0.109. Furthermore, it is worth noting that our method requires much lower computational parameters while delivering the best overall performance in this task.