Multi-channel speech enhancement method

Abstract

The invention discloses a multi-channel speech enhancement method in which a norm F is used for presenting the dimension of a signal subspace. The method comprises the following steps of: step 1, acquiring a multi-channel speech signal y(t) with a noise via a microphone array which is composed of N microphones, and calculating the noisy speech cross-correlation matrix Ryy of the multi-channel speech signal y(t), wherein t represents a discrete time point; step 2, estimating an additive noise cross-correlation matrix by virtue of a noise estimation algorithm; step 3, estimating a pure speech cross-correlation matrix by virtue of the noisy speech cross-correlation matrix Ryy and the additive noise cross-correlation matrix; step 4, estimating the dimension of the signal subspace by virtue of the pure speech cross-correlation matrix; step 5, performing generalized eigenvalue decomposition, and obtaining a time-domain constraint linear signal estimator by combining the dimension of the signal subspace and a Lagrangian multiplier mu; and step 6, filtering the multi-channel speech signal y(t) by virtue of the time-domain constraint linear signal estimator to obtain the enhanced speech.

Description

technical field [0001] The invention relates to the field of speech signal processing of speech enhancement and noise elimination, in particular to a multi-channel speech enhancement method. Background technique [0002] Signal Subspace Approach (SSA) is mainly based on matrix orthogonal decomposition theory, using Eigenvalue Decomposition (Eigenvalue Decomposition, EVD or Singular Value Decomposition, SVD) to decompose the vector space of noisy speech signal into Two parts that are orthogonal to each other: signal subspace (signal + noise) and noise subspace (noise only). Among them, EVD uses the autocorrelation matrix of the signal to realize the transformation from the time domain to the feature domain, which is regarded as dependent on the signal , which is often called the Karhunen-Loeve transform (KLT). Essentially, SSA-based speech enhancement is to zero the noise subspace while removing noise components in the signal subspace. [0003] SSA was first introduced into ...

AI Technical Summary

Problems solved by technology

However, in practical applications, due to the influence of factors such as room echo and non-stationary noise, the dimension of the signal subspace is often difficult to estimate accurately, so the prior knowledge of the low-rank model is difficult to accurately use

[0007] The fundamental defect of the traditional method of signal subspace dimension estimation is that it relies too much on noise estimation, while ignoring the variation law of the signal subspace dimension itself

When the SSA-based speech enhancement algorithm works in a real environment, due to the interference of noise, SVD or EVD will generate a large number of singular values ​​or eigenvalues ​​close to zero, and their value changes are continuous, which is also the signal subspace dimension. sure increased difficulty

Images