Effectiveness of Fine Linear Frequency Spectral Feature for Acoustic Event Detection
Effectiveness of Fine Linear Frequency Spectral Feature for Acoustic Event Detection
复制标题
精细线性频谱特征在声学事件检测中的有效性
DOI：
10.1109/gcce50665.2020.9291954
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Seiichi Nakagawa
中科院分区：
文献类型：
--
作者：
Kazumasa Yamamoto;Ryo Yamamoto;Seiichi Nakagawa
mel滤波器组特征通常用于声事件检测。然而，由于特征在较高频段的频率分辨率较低，因此没有利用较高的频率信息进行检测。本文采用精细线性频谱特征对DCASE2019声事件进行检测。我们改变了DCASE2019 Task 4基线的卷积神经网络结构以适应特征，并得到了对某些声学事件的检测改进。通过结合Mel-filterbank特征在分数水平上获得了对几乎声学事件的改进。
Mel-filterbank features are commonly used for acoustic event detection. However, due to the lower frequency resolution of the feature at higher frequency bands, the higher frequency information is not utilized for the detection. In this paper, we used fine linear frequency spectral features for the DCASE2019 acoustic event detection. We changed the convolutional neural network structure of the DCASE2019 Task 4 baseline to suite the features and got the improvement of detection for certain acoustic events. The improvement for almost acoustic events was obtained by combining with the Mel-filterbank features at the score level.