Page 207 - 《应用声学)》2023年第5期
P. 207

第 42 卷 第 5 期                                                                       Vol. 42, No. 5
             2023 年 9 月                          Journal of Applied Acoustics                 September, 2023

             ⋄ 研究报告 ⋄



                  一种基于聚类的门控卷积网络语声分离方法                                                                ∗






                                             罗 宇 胡维平            †    吴华楠


                                             (广西师范大学电子工程学院         桂林  541000)

                摘要:基于深度聚类的语声分离方法已被证明能有效地解决混合语声中说话人输出标签排列的问题,然而,现
                有关于聚类进行说话人分离方法,大多数是优化嵌入使每个源的重建误差最小化。该文以时域卷积网络为基
                础网络设计了一种改进基于聚类的门控卷积语声分离方法,在时域上通过堆叠的门控卷积网络,实现端到端
                深度聚类的源分离。该框架将非线性门控激活用于时域卷积网络中,提取语声信号的深层次特征;同时在高维
                特征空间中聚类对语声信号的特征进行表示和划分,为恢复不同信号源提供了一个长期的说话者表示信息。
                该框架解决了说话人输出标签排列问题并对语声信号的长期依赖性进行建模。通过华尔街日报数据集进行实
                验得出,该方法在信号失真比和尺度不变信号噪声比指标上分别达到了 16.72 dB 和 16.33 dB 的效果。
                关键词:深度聚类;门控卷积;语声分离
                中图法分类号: TN912.3           文献标识码: A          文章编号: 1000-310X(2023)05-1099-07
                DOI: 10.11684/j.issn.1000-310X.2023.05.024



               Clustering-based speech separation method for gated convolutional networks




                                           LUO Yu    HU Weiping    WU Huanan

                                 (Electronic Engineering, Guangxi Normal University, Guilin 541000, China)

                 Abstract: Deep clustering-based speech separation methods have been shown to be effective in solving the
                 problem of speaker output label alignment in mixed speech, however, most of the existing methods on clustering
                 for speaker separation optimize the embedding to minimize the reconstruction error of each source. In this
                 paper, we design an improved gate-convolutional cluster speech separation method based on the time-domain
                 convolutional network as the base network. The framework uses nonlinear gated activation in time-domain
                 convolutional networks to extract deep features of speech signals; and clustering in a high-dimensional feature
                 space to represent and segment the features of speech signals, providing a long-term speaker representation
                 information for recovering different sources. The framework solves the speaker output label alignment problem
                 and models the long-term dependency of speech signals. Experiments with the Wall Street Journal dataset yield
                 that the method achieves 16.72 dB and 16.33 dB in the signal distortion ratio and scale invariant signal-to-noise
                 ratio metrics, respectively.
                 Keywords: Deep clustering; Gated convolution; Speech separation





             2022-08-16 收稿; 2022-09-20 定稿
             国家自然科学基金项目 (NSFC 61861005)
             ∗
             作者简介: 罗宇 (1999– ), 男, 江西吉安人, 硕士研究生, 研究方向: 语声信号处理。
             † 通信作者 E-mail: huwp@gxnu.edu.cn
   202   203   204   205   206   207   208   209   210   211   212