基于深度压缩感知的语音增强模型

Speech enhancement model based on deep compressed sensing

随着压缩感知的深入研究,压缩感知在语音增强方面的应用也备受关注。针对传统压缩感知语音增强算法中存在的不足,将压缩感知与深度学习结合构建名为基于深度压缩感知的语音增强模型(Speech Enhancement based on Deep Compressed Sensing, SEDCS)。基于压缩感知原理使用编解码模型代替压缩感知中语音信号稀疏过程,使用卷积神经网络代替测量矩阵实现语音信号观测降维过程,通过联合训练的方式实现语音增强。实验结果表明:该模型能够完成语音增强任务,并且与现有的压缩感知语音增强算法相比,该模型能取得较好的语音增强效果;相比利用深度学习的语音增强算法,该模型虽性能一般,但在模型泛化性能和测试阶段的增强时间效率上有一定提升。

With the further research of compressed sensing, the application of compressed sensing in speech enhancement has attracted much attention. Aiming at the shortcomings of traditional compressed sensing speech enhancement algorithms, a speech enhancement model based on deep compressed sensing(SEDCS) is built by combining compressed sensing and deep learning. Based on the principle of compressed sensing, the codec model is used to replace the sparse process of speech in compressed sensing, and the convolutional neural network is used to replace the measurement matrix to realize the measurement and dimension reduction of speech. The speech enhancement of the model is obtained by jointly training. The experimental results show that the proposed model can complete the speech enhancement task and achieve good speech enhancement effect compared with the existing compressed sensing speech enhancement algorithm. Compared with the speech enhancement algorithm using deep learning, the performance of the model is general, but it is improved in the model generalization ability and the enhancement time efficiency in the test stage.