Nonnegative Matrix Factorization Based Adaptive Noise Sensing over Wireless Sensor Networks

Abstract

An adaptive noise sensing method is proposed to improve the speech sensing performance of speech-based applications operated over wireless sensor networks. The proposed method is based on nonnegative matrix factorization (NMF), which consists of adaptive noise sensing and noise reduction. In other words, adaptive noise sensing is performed by adapting a priori noise basis matrix of the NMF, which is estimated from the noise signal, resulting in an adapted noise basis matrix. Subsequently, the adapted noise basis matrix is used for the NMF decomposition of noisy speech into clean speech and background noise. The estimated clean speech signal is then applied to a front-end of the speech-based applications. The performance of the proposed NMF-based noise sensing and reduction method is first evaluated by measuring the source to distortion ratio (SDR), the source to interferences ratio (SIR), and the source to artifacts ratio (SAR). In addition, the proposed method is applied to an automatic speech recognition (ASR) system, which is a typical speech-based application, and then the average word error rate (WER) of the ASR is compared with that employing either a Wiener filter, or a conventional NMF-based noise reduction method using only a priori noise basis matrix.

Local Sparsity Based Online Dictionary Learning for Environment-Adaptive Speech Enhancement with Nonnegative Matrix Factorization

“…Dictionary Learning Figure 1 shows the procedure of the proposed speech enhancement method based on local sparsity estimation and online noise dictionary learning. Like the conventional NMF-based speech enhancement methods [8,9,13], the proposed method first decomposes spectral magnitude of noisy speech signal at the i-th frame, , into those of speech and noise, and , by using the supervised sparse NMF technique [13] with a fixed speech dictionary, , and an adaptive noise dictionary, ; . Subsequently, the local sparsity is calculated at every frequency bin by using a ratio between and , and then it is plugged into constructing an MMSE filter for both speech enhancement and online noise dictionary learning for the (i+1)-th frame.…”

“…where , , and denote the k-th spectral components of , , and , respectively. To separate and from , the p-powered spectral magnitude of noisy speech frame is represented as | | ≅ | | | | , according to satisfactory results of NMF-based noise reduction when is 1 or 2 [7][8][9][10][11][12][13][14]. For simplicity, | | , | | , and | | are represented as , , and , which are all 1 matrices.…”

“…In order to enhance speech signals recorded under such noise conditions, noise spectral components should be suppressed without damaging spectral components belonging to the target speech signal [3]. To this end, conventional efforts have been focused on estimating noise power spectra from the noisy signal [4,5] or separating speech and noise from the noisy speech [6][7][8][9][10][11][12][13][14][15] Among the successful noise estimation methods is the minima-controlled recursive algorithm (MCRA) [4,5]. MCRA estimates noise by only tracking minimum statistics in noise-only regions.…”

“…As an alternative to the minima tracking based noise estimation approaches, NMF-based source separation has been successfully realized into speech enhancement [7][8][9][10][11][12][13][14][15]. Several studies reported that NMF was suitable for separating speech signals from interfering non-stationary noise such as wind or television noise [7,8]. In general, the separation performance of the NMF-based methods is guaranteed if speech and noise NMF dictionaries are trained sufficiently to represent arbitrary noisy spectral magnitudes [7][8][9].…”

“…Several studies reported that NMF was suitable for separating speech signals from interfering non-stationary noise such as wind or television noise [7,8]. In general, the separation performance of the NMF-based methods is guaranteed if speech and noise NMF dictionaries are trained sufficiently to represent arbitrary noisy spectral magnitudes [7][8][9]. Thus, many NMF-based speech enhancement methods have trained speech and noise dictionaries in advance by using extensive speech and noise databases, respectively.…”