Supervised and semi-supervised speech enhancement using weighted nonnegative matrix factorization

According to the prior knowledge used in the algorithms, NMF-based speech enhancement can be categorized into supervised (NMF) one and semi-supervised (SNMF) one. In the supervised version, speech is estimated with the prior knowledge of both speech and noise. For the semi-supervised version, speech bases are learned on clean speech data in advance and adapted to different noise conditions. With the probabilistic estimation of whether the speech is present or not in a certain frame, this paper presents an extension to both NMF- and SNMF-based speech enhancement techniques by incorporating the speech presence probability (SPP) into their multiplicative update process. The performance of the algorithms is evaluated on three metrics (PESQ, LSD and SDR) by making experiments on TIMIT with 20 noise types at various signal-to-noise ratio (SNR) levels. All the proposed algorithms show the superiority over the conventional NMF, SNMF-based speech enhancement techniques, and also obtain better performance than some unsupervised state-of-the-art algorithms.