基于去噪自编码器和长短时记忆网络的语音测谎算法

为进一步提升语音测谎性能，提出了一种基于去噪自编码器（DAE）和长短时记忆（LSTM）网络的语音测谎算法。首先，该算法构建了优化后的DAE和LSTM的并行结构PDL；然后，提取出语音中的人工特征并输入DAE以获取更具鲁棒性的特征，同时，将语音加窗分帧后提取出的Mel谱逐帧输入到LSTM进行帧级深度特征的学习；最后，将这两种特征通过全连接层及批归一化处理后实现融合，使用softmax分类器进行谎言识别。CSC（Columbia-SRI-Colorado）库和自建语料库上的实验结果显示，融合特征分类的识别准确率分别为65.18%和68.04%，相比其他对比算法的识别准确率最高分别提升了5.56%和7.22%，表明所提算法可以有效提高谎言识别精度。

Speech deception detection algorithm based on denoising autoencoder and long short-term memory network

In order to further improve the performance of speech deception detection， a speech deception detection algorithm based on Denoising AutoEncoder （DAE） and Long Short-Term Memory （LSTM） network was proposed. Firstly， a parallel structure of DAE and LSTM was constructed， namely PDL （Parallel connection of DAE and LSTM）. Then， artificial features in the speech were extracted and put into the DAE to obtain more robust features. Simultaneously， the Mel spectrums extracted after adding windows to the speech and framing were input into LSTM frame-by-frame for frame-level depth feature learning. Finally， these two types of features were merged by the fully connected layer and the batch normalization， and the softmax classifier was used for the deception recognition. The experimental results on the CSC （Columbia-SRI-Colorado） corpus and the self-built corpus show that the recognition accuracy of the classification with fusion feature is 65.18% and 68.04% respectively， which is up to 5.56% and 7.22% higher than those of other algorithms， indicating that the proposed algorithm can effectively improve the accuracy of deception recognition.