An Unsupervised Speech Enhancement Method Based on Robust Nonnegative Matrix Factorization and Data Fusion

Abstract

The invention discloses a non-supervision speech enhancement method based robust non-negative matrix decomposition and data fusion. The method comprises windowing and framing input time-domain signals, and then performing Fourier transform and delivery to obtain the amplitude value of a sentence; performing estimation to obtain sparse phonetic components S and a noise basis matrix W (n); estimating speech components and noise components to obtain estimates of enhanced speech; decomposing the obtained estimates through a robust non-negative matrix, and fusing the decomposed estimates with estimates obtained through spectrum substraction (SS) and minimum mean square error (MMSE) through a geometric mean value filtering module to obtain final amplitude spectrum estimates; reconstituting the time-domain signals during speech enhancement through the amplitude spectrum estimates and the phases of noisy speeches. The non-supervision speech enhancement method based robust non-negative matrix decomposition and data fusion is not limited to the language that speech content belongs to, the change of speakers nor the variety of noise, and compared with traditional spectrum estimation algorithms of SS and MMSE which are based on stationarity assumption, can be free from dependence on the stationarity assumption and accurately estimate the frequency spectrum of stationary or mutational noise.

Description

technical field [0001] The invention belongs to the field of speech signal processing, in particular to an unsupervised speech enhancement method based on robust non-negative matrix decomposition and data fusion. Background technique [0002] Speech enhancement is of great significance both for improving the auditory effect of the speech signal and as a front-end processing to improve the performance of the speech recognizer. A key issue in the implementation of speech enhancement is noise estimation. In order to estimate the noise spectrum, some classical algorithms have been proposed, such as Spectrum Subtraction (SS), Minimum Mean Square Error (MMSE), etc., and have been widely used in voice communication. However, these methods are generally based on the assumption of stationarity of the noise, which is poor for spectral estimation of non-stationary abrupt noise. [0003] In order to estimate the spectrum of abrupt noise, noise estimation models based on dictionary lea...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the embodiments of the present invention is to provide an unsupervised speech enhancement method based on robust non-negative matrix decomposition and data fusion, which aims to solve the problem that existing supervised speech enhancement algorithms are limited to the language to which the speech content belongs, limited by The problem that speaker variation is limited by the type of noise

Images