基于张量分解的多声道音频恢复方法

Multi-Channel Audio Recovery Based on Tensor Decomposition

为恢复多声道音频在采集过程中丢失的数据,提出基于加权优化的张量分解方法.首先用张量对音频建模,并且根据其尺寸定义一个标识数据丢失位置的加权张量,然后使用加权最小二乘问题描述CANDECOMP/PARAFAC(CP)模型并通过一阶优化算法求解,最终通过获得的因子矩阵恢复音频.通过不同数目通道数据丢失的隐藏参考和基准的多激励测试,说明针对丢失数据采用CP分解方法是有效的,即张量分解能够得到较好的音频恢复效果.

In order to recover the missing data of multi-channel audio during collection,an efficient tensor decomposition method was proposed based on weight optimization.Firstly,multi-channel audio was represented as an audio tensor.A weight tensor of the identical size as audio tensor was defined,and which identified the location of missing channel data.Afterward,the CANDECOMP/PARAFAC decomposition（CPD）was formulated as a weighted least squares problem which was solved by using a first-order optimization approach.At last,the audio were recovered via achieved factor matrices.In experiments about varying number of missing channel entries,the results of multiple stimuli with hidden reference and anchor show that it is validated by CPD of multi-channel audio in the presence of missing data,and tensor decomposition is a successful approach for recover audio.