摘要
为了使盲源分离算法能更好地应用于一些实际噪声和训练数据较少且不需要做标记的环境,文章提出了一种无监督的非负矩阵字典学习方法。该方法对混合信号进行字典学习,随后在每个时间点上根据其空间源对字典原子进行分组来实现分离。通过从SiSEC获取语音和现实噪声的两通道混合信号作为数据集,使用PEASS和BSS Eval工具包分别基于感知、基于SNR和PEMO-Q的度量来量化性能。此外,还评估模型了参数对分离质量的影响,并将该方法与其他无监督和半监督的语音分离方法进行比较。结果证明,GCC-NMF是一种灵活的源分离算法,在3种评估参数中的每个参数均胜过特定任务的方法,包括盲源以及需要先验知识或信息的多种已知方法。
An unsupervised non-negative matrix dictionary learning method is proposed to make the blind source separation algorithm better applicable under the circumstances with actual noise and less training data which don′t need to be marked.Dictionary learning is performed on the mixture signal and separated by grouping dictionary atoms according to their spatial origins.By acquiring a two-channel mixed signal of speech and real noise from SiSEC as a data set,the PEASS and BSS Eval toolkits are used to quantify performance using perceptual-based,SNR-based,and PEMO-Q metrics,respectively.Besides,the effect of separation quality via model parameters is also evaluated and compared with other unsupervised and semi-supervised separation methods.The results prove that GCC-NMF is a flexible algorithm for origins separation,as each parameter is superior to that from other specific target approaches,including blind separation speech and other existed approaches that require priori knowledge and information.
作者
吴君钦
王迎福
WU Junqin;WANG Yingfu(School of Information Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,Jiangxi,China)
出处
《江西理工大学学报》
CAS
2020年第5期65-72,共8页
Journal of Jiangxi University of Science and Technology
基金
国家自然科学基金资助项目(61741109)。
