Diffuse noise suppression with asynchronous microphone array based on amplitude additivity model
Diffuse noise suppression with asynchronous microphone array based on amplitude additivity model
复制标题
基于幅度相加模型的异步麦克风阵列漫噪声抑制
DOI：
10.1109/apsipa.2015.7415339
复制
发表时间：
2015
期刊：
影响因子：
--
通讯作者：
Takeshi Yamada and Shoji Makino
中科院分区：
文献类型：
--
作者：
Yoshikazu Murase;Hironobu Chiba;Nobutaka Ono;Shigeki Miyabe;Takeshi Yamada and Shoji Makino
本文提出了一种基于非负矩阵分解（NMF）及其有效半监督训练的多通道幅度分析方法，用于抑制大量干扰。为了减少异步麦克风阵列的点源干扰，我们提出了基于幅度的时信道域语音增强，我们称之为传递函数增益NMF。传递函数增益NMF是一种抗漂移的鲁棒方法，它会破坏信道间相位分析。我们使用这种方法来抑制大量的源。我们表明，假设噪声源离麦克风足够远，并且空间特征彼此相似，那么大量的干扰可以通过单一基来建模。由于NMF参数的盲优化不能很好地用于被持续的重噪声污染的稀疏观测，我们在噪声抑制之前使用语音缺失观测训练弥漫性噪声基，这可以通过简单的语音活动检测技术轻松获得。我们在模拟被漫射噪声包围的目标源的实验中证实了我们提出的模型和半监督传递函数增益NMF的有效性。
In this paper, we propose a method for suppressing a large number of interferences by using multichannel amplitude analysis based on nonnegative matrix factorization (NMF) and its effective semi-supervised training. For the point-source interference reduction of an asynchronous microphone array, we propose amplitude-based speech enhancement in the time-channel domain, which we call transfer-function-gain NMF. Transfer-function-gain NMF is a robust method against drift, which disrupts an inter-channel phase analysis. We use this method to suppress a large number of sources. We show that a mass of interferences can be modeled by a single basis assuming that the noise sources are sufficiently far from the microphones and the spatial characteristics become similar to each other. Since the blind optimization of the NMF parameters does not work well with merely sparse observation contaminated by the constant heavy noise, we train the diffuse noise basis in advance of the noise suppression using a speech absent observation, which can be obtained easily using a simple voice activity detection technique. We confirmed the effectiveness of our proposed model and semi-supervised transfer-function-gain NMF in an experiment simulating a target source that was surrounded by a diffuse noise.