基于注意力机制的NewVGG16-BiGRU鼾声分类

NewVGG16-BiGRU snoring classification based on attention mechanism

针对已有的鼾声分类模型因未考虑实际睡眠时的其他声音而导致的泛化能力差、准确率较低等问题,提出一种基于注意力机制的NewVGG16双向门控循环单元(NVGG16-BiGRU-Att)算法用于鼾声识别。首先,生成每个声段的谱图,采用NVGG16网络提取语谱图、梅尔(Mel)时频图和恒Q变换(CQT)时频图组成的谱图特征矩阵;其次,将提取的特征向量输入BiGRU,结合注意力机制,增加分类过程中的重要特征信息的权重,改善分类效果;最后,经过全连接层输出鼾声与非鼾声。在采集的鼾声数据集上进行实验,实验结果表明,所提算法取得了较好的分类效果,其中Mel时频图效果最优,识别准确率达到96.18%;相较于卷积神经网络(CNN)+长短期记忆(LSTM)网络、卷积CNNsLSTMs-深度神经网络(DNNs)模型,在同特征输入下,所提算法的准确率提升了0.31%~2.39%,验证了所提算法具有较好的鲁棒性,能够提升分类性能。

Aiming at the poor generalization ability and low accuracy of existing snoring classification models due to failure to consider other sounds during sleep,a NewVGG16 Bidirectional Gated Recurrent Unit based on Attention mechanism(NVGG16-BiGRU-Att)algorithm was proposed for snoring recognition.Firstly,the spectrogram of each sound segment was generated,and NVGG16 network was used to extract the spectrogram feature matrix composed of spectrogram,Mel time-frequency map and Constant Q Transform(CQT)time-frequency map.Secondly,the extracted feature vector was input into the BiGRU,combined with the attention mechanism,the weight of important feature information in the classification process was enhanced and the classification effect was improved.Finally,the snoring and non-snoring were output through the fully connected layer.Experiments were launched on the collected snoring dataset.Experimental results show that the proposed algorithm achieves a good classification effect.Among them,Mel time-frequency map has the best effect,and the recognition accuracy reaches 96.18%.Compared with Convolutional Neural Network(CNN)+Long Short Term Memory(LSTM)network and CNNs-LSTMs-Deep Neural Networks(DNNs)models,under the same feature input,the accuracy of the proposed algorithm is improved by 0.31%-2.39%,which verifies that the proposed algorithm has better robustness and can improve classification performance.