Online parameter selection for source separation using non-negative matrix factorization

Non-negative matrix factorization;source separation;spectral subtraction;noise removal;speech enhancement

Kang, Kang ; Smaragdis ; Paris

关键词: Non-negative matrix factorization;    source separation;    spectral subtraction;    noise removal;    speech enhancement;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/34281/Kang_Kang.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Blind source separation has been an area of study recently due to the many applications that might benefit from a good blind source separation algo- rithm. One instance is using blind source separation for audio denoising in cellular phones. In almost all instances, we have very little, if any, infor- mation about how background noise is mixed with the speaker’s voice in a given cell phone conversation. Current techniques include spectral subtrac- tion and Wiener filtering which are classical DSP techniques to deal with stationary noises. In this document, we aim to present a study on how to use blind source separation algorithms to denoise audio mixtures containing speech and various background noises. We mainly focus on how to imple- ment an online source separation algorithm which can handle non-stationary noises. To address the implementation, we also present a study on how to select the parameters in the separation algorithm in order to deliver the best performance for denoising using a statistical metric we have defined.