CIAIC-BAD SYSTEM FOR DCASE 2018 CHALLENGE TASK 3 Technical Report

In this technical report, we present our system for the task 3 of Detection and Classification of Acoustic Scenes and Events 2018 (DCASE2018) challenge, i.e. bird audio detection(BAD). First, log mel-spectrogram and mel-frequency cepstral coefficients (MFCC) are extracted as features. In order to improve the quality of original audio, same denoising methods are adopted, for example, adaptive denoising in Adobe Audition. Then, convolutional recurrent neural networks (CRNN) with customized activation function is used for detection. Finally, we use aforementioned features as inputs to train our CRNN model and make a fusion on three subsystems to further improve the performance. We evaluate the proposed systems on the dataset with area under the ROC curve (AUC) measure, and our best AUC score on leaderboard dataset is 85.67.