基于CNN和LSTM融合特征提取的车内声品质评价模型研究

基于深度学习方法建立的车内声品质评价模型不需要高度依赖声学理论和经验知识，可以有效提取深层次特征，客观高效地获得符合主观感受的评价结果。为获取噪声中符合人耳对声音感受的频率信息，便于在深度学习中进行特征提取，采用对数梅尔频谱和时频遮掩相结合的方法对采集到的噪声样本进行预处理。为有效提取车内噪声深层次特征，融合卷积神经网络 （Convolution Neural Network，CNN） 和长短时记忆网络 （Long Short-Term Memory Network，LSTM） 各自的优点，建立了融合特征提取层。使用全连接和Softmax输出单元组合构建了分类器模块。在合适的超参数下，模型通过充足的训练获得了96.88%的训练准确度。使用大量样本对模型进行验证，得到93.69%的验证准确度；采用混淆矩阵对模型进一步验证，总体的预测评价等级与真实评价等级偏差不大，证明模型的预测结果与主观评价结果具有很好的一致性。

Research on Sound Quality Evaluation of Vehicle Interior Noise Based on Merged Features Extracted by CNN and LSTM

The evaluation model of vehicle interior sound quality based on deep learning does not rely heavily on acoustic theory and empirical knowledge, can effectively extract deep-level features and efficiently obtains evaluation results in accordance with subjective feelings. Initially the collected noise samples was preprocessed by using a method combining logarithmic Mel spectrum and time-frequency masking to obtain the samples in the range of human frequency perception for feature extraction in deep learning. And then a fusion feature extraction layer combining CNN and LSTM was established to effectively extract the deeplevel features of vehicle noise. A classifier module was built which consisted a fully connected layer and the Softmax output unit. With suitable hyperparameters, the model achieved a training accuracy of 96.88% after sufficient training. Finally a large number of test samples were used to verify the model and the verification accuracy was 93.69%. The confusion matrix was used for further validation. And there is little difference between the overall predicted evaluation levels and the real evaluation levels; i. e., the prediction results of vehicle interior sound quality are in good agreement with the subjective evaluation results.