Reducing the interference noise in a monaural noisy speech signal has been achallenging task for many years. Compared to traditional unsupervised speechenhancement methods, e.g., Wiener filtering, supervised approaches, such asalgorithms based on hidden Markov models (HMM), lead to higher-quality enhancedspeech signals. However, the main practical difficulty of these approaches isthat for each noise type a model is required to be trained a priori. In thispaper, we investigate a new class of supervised speech denoising algorithmsusing nonnegative matrix factorization (NMF). We propose a novel speechenhancement method that is based on a Bayesian formulation of NMF (BNMF). Tocircumvent the mismatch problem between the training and testing stages, wepropose two solutions. First, we use an HMM in combination with BNMF (BNMF-HMM)to derive a minimum mean square error (MMSE) estimator for the speech signalwith no information about the underlying noise type. Second, we suggest ascheme to learn the required noise BNMF model online, which is then used todevelop an unsupervised speech enhancement system. Extensive experiments arecarried out to investigate the performance of the proposed methods underdifferent conditions. Moreover, we compare the performance of the developedalgorithms with state-of-the-art speech enhancement schemes using variousobjective measures. Our simulations show that the proposed BNMF-based methodsoutperform the competing algorithms substantially.
展开▼
