Speech enhancement via ensemble modeling NMF adaptation
Jeremy Chiaming Yang, Syu‐Siang Wang, Yu Tsao, Jeih-weih Hung · 2016
Nonnegative matrix factorization (NMF)-based speech enhancement algorithm has been proven to provide satisfactory performance when the prior information about speaker and noise types are given. In most real-world scenarios, however, such prior information may not always be accessible. Therefore, an adaptation technique is favorable to adapt the NMF matrices to match the testing condition for a better enhancement performance. In this study, we proposed a novel ensemble modeling (ESM) algorithm for NMF adaptation. The algorithm effectively uses the local information of the entire training data to facilitate an effective and efficient online NMF adaptation. The adapted NMF matrices were used to perform speech enhancement. Experimental results on the perceptual evaluation of speech quality (PESQ) confirmed the effectiveness of the ESM algorithm in various signal-to-noise ratio conditions.