【摘要】 深度神经网络声音事件检测方法需要大量标记声音事件类别和起止时间的强标签音频样本，然而强标签标注非常困难和耗时.弱标签声音事件检测是解决这一困难的有效途径.本文将弱标签声音事件检测作为多实例学习问题，并基于卷积循环神经网络提出弱标签声音事件检测的空间-通道特征表征与自注意池化方法 .该方法研究多实例弱标签声音事件检测的特征表征和帧级预测结果池化两个方面的内容.在特征表征方面，为了增强卷积神经网络的特征表征能力，结合上下文门控和通道注意机制构建门控注意力结构并嵌入到卷积循环神经网络中，实现了音频样本特征的空间和通道特征选择；在预测结果池化方面，引入自注意思想设计音频帧预测结果的自注意池化方法，增强了音频样本中事件帧之间的相关度，使事件帧获得更大的权重.本文方法通过对卷积循环神经网络特征表征和预测结果池化的革新，有效提升了模型的检测性能.本文提出的方法在DCASE 2017任务4和DCASE 2018任务4数据集的评估集中分别取得了52.47%和31.00%的F1得分，性能优于当前绝大部分的弱标签声音事件检测方法 .实验结果表明：本文提出的空间-通道特征表征与自注意池化方法能显著改善弱标签声音事件检测的综合性能.更多还原

【Abstract】 A large amount of strong labeled audio samples, which are annotated with detailed sound event categories and timestamps, is required for a deep neural network sound event detection(SED) model. However, obtaining strong label is very difficult and time-consuming. Weakly-labeled SED is an effective way to solve this problem. This paper approaches weakly-labeled SED as a multiple instance learning(MIL) problem and proposes a spatial-channel feature representation and self-attention pooling method for weakly-labeled SED based on convolutional recurrent neural network(CRNN). The proposed method studies the feature representation and the frame-level prediction pooling method for multi-instance weaklylabeled SED. In feature representation, in order to enhance the ability of CRNN, we design a gating attention structure by combining context gating and channel attention mechanism, and embed it into CRNN to realize the spatial and channel selection of audio sample features. In frame-level prediction pooling, we introduce the idea of self-attention and design a selfattention pooling(SAP) function to enhance the event frame correlation in the audio sample and assign great weights for event frames. The proposed method effectively improves the detection performance of SED model by innovating the feature representation of CRNN and the pooling method of frame-level predictions. The proposed method has achieved 52.47% and 31.00% F1 scores respectively in the evaluation set of DCASE 2017 task 4 and DCASE 2018 task 4 datasets, which outperforms most of the current weakly-labeled SED methods. Experimental results show that the proposed spatial-channel feature representation and self-attention pooling method can significantly improve the performance of weakly-labeled SED.更多还原