Speech enhancement based on nonnegative matrix factorization with mixed group sparsity constraint

Abstract

International audienceThis paper addresses a challenging single-channel speech enhancement problem in real-world environment where speech signal is corrupted by high level background noise. While most state-of-the-art algorithms tries to estimate noise spectral power and filter it from the observed one to obtain enhanced speech, the paper discloses another approach inspired from audio source separation technique. In the considered method, generic spectral characteristics of speech and noise are first learned from various training signals by non-negative matrix factorization (NMF). They are then used to guide the similar factorization of the observed power spectrogram into speech part and noise part. Additionally, we propose to combine two existing group sparsity-inducing penalties in the optimization process and adapt the corresponding algorithm for parameter estimation based on mul-tiplicative update (MU) rule. Experiment results over different settings confirm the effectiveness of the proposed approach

Gaussian Modeling-Based Multichannel Audio Source Separation Exploiting Generic Source Spectral Model

“…We will show in the experiment that this way of constructing the GSSM does not provide as good source separation performance as the one presented before by (18).…”

“…where U j is constructed by (18) or (20) and fixed, Ω( H j ) presents a penalty function imposing sparsity on H j , and λ is a trade-off parameter determining the contribution of the penalty. Note that as the GSSM U j constructed in (18) becomes a large matrix when the number of examples L j for each source increases, and it is actually a redundant dictionary since different examples may share similar spectral patterns.…”

“…Inspired by this idea, we have further investigated the use of generic speech and noise model for single-channel speech separation in [16] and shown its promising result in (a) the supervised case, where both speech GSSM and noise GSSM are learned during training phase, and (b) the semi-supervised case, where only the speech GSSM is pre-learned. Furthermore, we have proposed to combine the block sparsity constraint investigated in [14] with the component sparsity constraint presented in [17] in a common formulation in order to take into account the advantage of both of them [18].…”

“…It should be noted that the works cited above [9], [12], [16], [18] considered only a single channel case, where the mixtures are mono, and exploited non-negative matrix factorization (NMF) [19], [20] to model the spectral characteristics of audio sources. Some recent works have investigated the use of the deep neural networks (DNN) to model the source spectra, where basically the types of sources in the mixture also need to be known as a side information in order to collect training data.…”

“…In this paper, we present an extension of the previous works [15], [16], [18] to the multichannel case where the NMF-based GSSM is combined with the full-rank spatial covariance model in a Gaussian modeling paradigm. Around this LGM, existing works have investigated several source spectral models such as Gaussian mixture model (GMM) [37], NMF as a linear model with nonnegativity constraints [36], [38], continuity model [39], kernel additive model [40], heavy-tailed distributionsbased model [41], [42], and recently DNN [24].…”