Column-wise update algorithm for independent deeply learned matrix analysis
Column-wise update algorithm for independent deeply learned matrix analysis
复制标题
DOI：
--
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Yu Takahashi
中科院分区：
文献类型：
--
作者：
Yu Takahashi
本文提出了一种用于音频源分离的鲁棒除混滤波器更新算法。音频源分离是一项从麦克风阵列中观察到的多通道混合中恢复源信号的任务，可应用于语音识别和音乐信号分析等。近年来，独立深度学习矩阵分析（IDLMA）作为一种最先进的分离方法被提出。IDLMA利用源模型的深度神经网络（DNN）推理和基于源独立性的去混滤波器的盲估计。在传统的IDLMA中，使用迭代投影（IP）来估计除混滤波器。虽然IP是一种快速算法，但当特定源模型由于信噪比较差而不准确时，后续的滤波器更新会失败。这是因为IP以源方式更新去混过滤器，每次更新只使用一个源模型。在本文中，我们推导了一种新的基于麦克风的更新规则，该规则在每次更新时同时利用源模型的所有信息。此外，我们提出了一种方法，根据通过dnn估计的源信号的伪信噪比来选择合适的源或麦克风更新规则。实验结果表明了该方法的有效性。
In this paper, we propose a robust demixing filter update algorithm for audio source separation. Audio source separation is a task to recover source signals from multichannel mixtures observed in a microphone array, which can be applied to, e.g., speech recognition and music signal analysis. Recently, independent deeply learned matrix analysis (IDLMA) has been proposed as a state-of-the-art separation method. IDLMA utilizes deep neural network (DNN) inference of source models and blind estimation of demixing filters based on sources’ independence. In conventional IDLMA, iterative projection (IP) is exploited to estimate the demixing filters. Although IP is a fast algorithm, when the specific source model is not accurate owing to the bad SNR condition, the successive update of filters will fail hereafter. This is because IP updates the demixing filters in a source-wise manner where only one source model is used for each update. In this paper, we derive a new microphone-wise update rule which exploits all information of the source models simultaneously for each update. Moreover, we propose a method to select the appropriate sourceor microphone-wise update rule depending on the source signal’s pseudo SNR estimated via the DNNs. Experimental results show the efficacy of the proposed method.