• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.479-487
• /
• 2021

In this paper, we propose an Sound Event Detection (SED) model using self-training based on a noisy student model. The proposed SED model consists of two stages. In the first stage, a mean-teacher model based on an Residual Convolutional Recurrent Neural Network (RCRNN) is constructed to provide target labels regarding weakly labeled or unlabeled data. In the second stage, a self-training-based noisy student model is constructed by applying different noise types. That is, feature noises, such as time-frequency shift, mixup, SpecAugment, and dropout-based model noise are used here. In addition, a semi-supervised loss function is applied to train the noisy student model, which acts as label noise injection. The performance of the proposed SED model is evaluated on the validation set of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2020 Challenge Task 4. The experiments show that the single model and ensemble model of the proposed SED based on the noisy student model improve F1-score by 4.6 % and 3.4 % compared to the top-ranked model in DCASE 2020 challenge Task 4, respectively.

• Journal of Broadcast Engineering
• /
• v.24 no.3
• /
• pp.472-484
• /
• 2019

This paper proposes a time-domain sound event detection algorithm using DNN (Deep Neural Network). In this system, time domain sound waveform data which is not converted into the frequency domain is used as input to the DNN. The overall structure uses CRNN structure, and GLU, ResNet, and Squeeze-and-excitation blocks are applied. And proposed structure uses structure that considers features extracted from several layers together. In addition, under the assumption that it is practically difficult to obtain training data with strong labels, this study conducted training using a small number of weakly labeled training data and a large number of unlabeled training data. To efficiently use a small number of training data, the training data applied data augmentation methods such as time stretching, pitch change, DRC (dynamic range compression), and block mixing. Unlabeled data was supplemented with insufficient training data by attaching a pseudo-label. In the case of using the neural network and the data augmentation method proposed in this paper, the sound event detection performance is improved by about 6 %(based on the f-score), compared with the case where the neural network of the CRNN structure is used by training in the conventional method.