Performance analysis of the convolutional recurrent neural network on acoustic event detection

In this study, we attempted to find the optimal hyper-parameters of the convolutional recurrent neural network (CRNN) by investigating its performance on acoustic event detection. Important hyper-parameters such as the input segment length, learning rate, and criterion for the convergence test, were determined experimentally. Additionally, the effects of batch normalization and dropout on the performance were measured experimentally to obtain their optimal combination. Further, we studied the effects of varying the batch data on every iteration during the training. From the experimental results using the TUT sound events synthetic 2016 database, we obtained optimal performance with a learning rate of . We found that a longer input segment length aided performance improvement, and batch normalization was far more effective than dropout. Finally, performance improvement was clearly observed by varying the starting points of the batch data for each iteration during the training.