功率比相关子带划分快速独立向量分析

Fast Independent Vector Analysis using Power Ratio Correlation-based Bands Partition

传统独立向量分析利用频点之间的高阶相关性解决盲源分离频域排序问题,已有研究表明,频点之间的高阶相关性与频点间距有关,越近的频点相关性越强。考虑此特点,本文提出在频域进行无重叠子带划分,采用功率比相关的方法解决子带之间的排序问题;结合更符合语音分布模型的多变量广义高斯分布和多变量t分布,实现了性能更优的功率比相关子带划分快速独立向量分析算法。实验结果表明,本文提出的算法相比传统独立向量分析算法具有更好的语音分离性能。

Traditional independent vector analysis resolves permutation ambiguity using the higher-order dependency among the whole frequency band.Researchers have shown that neighboring frequencies have stronger dependency and using frequency bands partition can improve the separation result.Firstly,in this work,the overlapping cliques independent vector analysis based on natural gradient was extended to a fast algorithm using Newton gradient.Secondly,multivariate generalized Gaussian distribution and multivariate Student t distribution were introduced as source distribution priors in overlapping cliques or overlapping bands partition fast independent vector analysis algorithms because they were more suitable to model the heavy-tailed property of speech signals.Finally,a non-overlapping bands partition scheme was proposed in the fast independent vector analysis with heavy-tailed distributions.Power ratio correlation was introduced to avoid the block permutation ambiguity between frequency bands.Both simulation and real recording experimental results show that the proposed algorithm is better than the traditional fast independent vector analysis and other overlapping bands partition algorithms.