| 利用奇异谱分析的深度神经网络语音增强方法 |
| |
| 引用本文： | 王骞,何培宇,徐自励.利用奇异谱分析的深度神经网络语音增强方法[J].信号处理,2020,36(6):902-910. |
| |
| 作者姓名： | 王骞 何培宇 徐自励 |
| |
| 作者单位： | 四川大学电子信息学院 |
| |
| 基金项目： | 国家自然科学基金项目(61071159, U1733109) |
| |
| 摘 要： | 针对现有深度神经网络语音增强方法对带噪语音的去噪能力有限、语音质量提升不高的问题,提出了一种基于奇异谱分析的深度神经网络语音增强方法。通过引入奇异谱分析算法对带噪语音进行预处理,以初步分离得到语音信号与噪声。接着将语音信号与噪声用于深度神经网络模型的训练,以得到性能更优的网络模型,从而使得本文方法具有更好的性能。最后在重建干净语音的环节中,同时使用神经网络估计得到的对数功率谱和带噪语音的对数功率谱,并加入了权重系数,使得本文提出的方法可以适应不同信噪比的情形,有效的去除背景噪声,降低语音信号的失真。本文通过仿真实验验证了该方法的有效性和鲁棒性。
|
| 关 键 词： | 奇异谱分析 深度神经网络 对数功率谱 语音增强 |
| 收稿时间： | 2020-03-04 |
Deep Neural Network speech Enhancement method based on singular Spectrum Analysis |
| |
| Wang Qian,He Peiyu,Xu Zili.Deep Neural Network speech Enhancement method based on singular Spectrum Analysis[J].Signal Processing,2020,36(6):902-910. |
| |
| Authors: | Wang Qian He Peiyu Xu Zili |
| |
| Affiliation: | College of Electronic Information, Sichuan University |
| |
| Abstract: | In order to solve the problems of limited denoising ability and low speech quality improvement of existing deep neural network speech enhancement methods for noisy speech, a deep neural network speech enhancement method based on singular spectrum analysis is proposed. By introducing the singular spectrum analysis algorithm to preprocess the noisy speech, the speech signal and noise are preliminarily separated. Then the speech signal and noise are used to train the depth neural network model to obtain a network model with better performance, so that the new method of deep neural network speech enhancement based on singular spectrum analysis has better performance. Finally, both the logarithmic power spectrum estimated by the neural network and the logarithmic power spectrum of noisy speech are used to reconstruct clean speech. The method proposed in this paper can adapt to the situation of different signal-to-noise ratio, effectively remove the background noise and reduce the distortion of speech signal. In this paper, simulation experiments are carried out to verify the effectiveness and robustness of the method. |
| |
| Keywords: | singular spectrum analysis depth neural network logarithmic power spectrum speech enhancement |
| 本文献已被 维普 等数据库收录！ |
| 点击此处可从《信号处理》浏览原始摘要信息 |
|
点击此处可从《信号处理》下载全文 |
|
