Regularized non-negative matrix factorization with Gaussian mixtures and masking model for speech enhancement

详细信息    查看全文

• 作者：Hanwook Chung ; ^a ; ^{hanwook.chung@mail.mcgill.ca} ; Eric Plourde^b ; ^{eric.plourde@usherbrooke.ca} ; Benoit Champagne^a ; ^{benoit.champagne@mcgill.ca}
• 关键词：Single-channel speech enhancement ; Supervised algorithm ; Non-negative matrix factorization ; Regularization ; Gaussian mixture model ; Masking effects
• 刊名：Speech Communication
• 出版年：2017
• 出版时间：March 2017
• 年：2017
• 卷：87
• 期：Complete
• 页码：18-30
• 全文大小：2576 K
• 卷排序：87

文摘

We propose regularized NMF algorithms for supervised single channel speech enhancement. The a priori magnitude spectral distributions are modeled by Gaussian mixtures. Corresponding log-likelihood functions are used as regularization terms. We construct a weighted Wiener filter based on the masking threshold.