EI / SCOPUS / CSCD 收录

中文核心期刊

利用深度学习的运动声源线谱增强方法

Moving sound source line spectrum enhancement method using deep learning

  • 摘要: 针对运动声源线谱存在的低信噪比、多普勒频移等情况, 提出了使用基于监督学习的深度学习方法实现线谱增强。训练数据为本地生成的干净和含噪线谱数据对。分别采用直接参数映射和比值掩码方法进行训练并建立线谱增强模型, 最后使用两种模型处理了仿真数据与SWellEx-96实验数据, 并将处理结果与传统的增强方法进行对比。对于仿真数据, 两种模型均可用于线谱增强, 且均具备加强多线谱和多普勒频移线谱的能力。当信噪比为−35 dB, 两种模型可分别实现约 11.74 dB信噪比增益和7.52 dB降噪。对于SWellEx-96实验数据, Masking模型对多普勒频移性质的线谱具有良好的不失真性质, 证明了所提深度学习模型应用于运动声源线谱增强的可行性。

     

    Abstract: Considering the low signal-to-noise ratio (SNR) and Doppler frequency shift of the line spectrum of a moving sound source, this paper proposes a deep learning method based on supervised learning to achieve line spectrum enhancement. The training datasets are locally generated clean and noisy line spectra, which are trained by mapping and masking methods, respectively. Finally, the simulation data and the SWellEx-96 experimental data are compared and verified with the traditional enhancement method. For the simulation data, both models can be used for line spectrum enhancement, and both models could enhance multi-line spectrum and line spectrum with Doppler frequency shift. When the input SNR is −35 dB, the two models can achieve about 11.74 dB SNR gain and 7.52 dB noise reduction, respectively. For the SWellEx-96 dataset, the masking model has undistorted property for the line spectra with Doppler shift, which proves the feasibility of applying the proposed deep learning model to enhance the line spectra of moving sound sources.

     

/

返回文章
返回