A Bayesian approach to convolutive nonnegative matrix factorization for blind speech dereverberation

Abstract

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. Highlights • A dereverberation method that needs no pre-processing nor specific information. • Very good performance observed in real-world recording conditions. • Performance speed is fast enough to use as startpoint for on-line dereverberation. • The theoretical ground is solid and allows for easily seeking ways of improvement.

Recognition of Audio Depression Based on Convolutional Neural Network and Generative Antagonism Network Model

“…Their method can mine the representative features by utilize the visual bag of words method to accelerate the extraction process of robust features of audio, and then quantify it into a set of visual words and image histograms. In [20] work, they proposed a bayesian audio recognition method based on convolutional non-negative matrix factorization, which imposed certain features on the time-frequency components of the recovered signal and reverberation component through the prior distribution. Hoirin and etl.…”

Switching Divergences for Spectral Learning in Blind Speech Dereverberation

“…with uniform distribution in [−π, π). Under this hypothesis, it can be shown ( [7]) that the expected value of |x k [t]| 2 is given by…”

“…From these characteristics, and the fact that the overlapping of windows results in consecutive time components of H capturing common information, it is reasonable to expect the components of H to exhibit a smooth decay over time ( [16]). This structure can be promoted (see [7]) by introducing a penalizer of the form…”

“…3. Test the accuracy of the representation Ŝ by computing the cepstral distance with respect to S. It should be pointed out that functional ( 5) is a generalization of a Bayesian approach (similar to the one in [7]) if U and ∇ t H are treated as random variables with exponential and normal a-priori distributions, respectively. In fact, by choosing β = 2, the minimizer of ( 5) corresponds to a maximum-a-posteriori (MAP) estimator, given proper choices of the penalization parameters.…”

“…Due to the characteristics of speech signals, most state-of-the-art methods deal with the dereverberation problem in a transformed domain, such as the one obtained by the Fan-Chirp Transform (see [6]) or the Short-Time Fourier Transform (STFT) ( [7]). Some of these methods make use of non-negative matrix factorization (NMF) or its variants, such as convolutive NMF ( [8]), along with Bayesian or penalization approaches.…”