Experimental Evaluation of Multichannel Audio Source Separation Based on IDLMA
Experimental Evaluation of Multichannel Audio Source Separation Based on IDLMA
复制标题
DOI：
--
复制
发表时间：
2018-03
期刊：
影响因子：
--
通讯作者：
Kitamura Daichi;Sumino Hayato;Takamune Norihiro;Takamichi Shinnosuke;Saruwatari Hiroshi;Nobutaka Ono
中科院分区：
文献类型：
--
作者：
Kitamura Daichi;Sumino Hayato;Takamune Norihiro;Takamichi Shinnosuke;Saruwatari Hiroshi;Nobutaka Ono
在本文中，我们提出了一种新的知情多通道音频源分离，称为独立深度学习矩阵分析（IDLMA）。 IDLMA是传统盲源分离、独立低秩矩阵分析和基于深度神经网络（DNN）的监督学习方法的统一算法，可以解释为基于独立的源分离理论的自然知情扩展。虽然源模型是通过预先训练的sourcewise DNN 来估计的，但空间模型可以通过源之间的统计独立性来盲目地估计。使用音乐信号的实验显示了 IDLMA 与传统的基于 DNN 的技术相比的有效性。
In this paper, we propose a new informed multichannel audio source separation called independent deeply learned matrix analysis (IDLMA). IDLMA is a unified algorithm of conventional blind source separation, independent low-rank matrix analysis, and a supervised learning method based on deep neural networks (DNN) and can be interpreted as a natural informed extension of the independence-based source separation theory. Although a source model is estimated by pre-trained sourcewise DNN, a spatial model can blindly be estimated by statistical independence between sources. The experiment using music signals shows the efficacy of IDLMA compared with the conventional DNN-based techniques.