Audio enhancement method and system

Abstract

The invention discloses an audio enhancement method. The spatial spectrum of the original multi-channel audio is obtained by the direction of arrival estimation algorithm. Acquiring multiple peaks greater than a set threshold from the spatial spectrum; acquiring multiple estimated direction values ​​of the multiple peaks according to a DOA wave-of-arrival estimation method. A spatial covariance matrix of multiple estimated direction values ​​is obtained according to the multiple estimated direction values ​​and the steering vector of the microphone array. The CGMM complex Gaussian mixture model is initialized and established according to the spatial covariance matrix; the parameters of the CGMM complex Gaussian mixture model are iteratively updated by the clustering method. Enhanced audio is obtained by enhancing the original multi-channel audio through the MVDR minimum variance distortion-free response beamforming algorithm. This method reduces the number of iterations of the EM algorithm to update the parameters of the CGMM model, and greatly reduces the amount of calculation. At the same time, the category of time-frequency point masking values ​​obtained in each frequency band is definite, so that the masking values ​​of the same category in each frequency band can be merged together, which overcomes the problem of fuzzy sorting.

Description

technical field [0001] The invention belongs to the technical field of speech recognition, in particular to an audio enhancement method and system. Background technique [0002] At present, the masking value of the time-frequency point is mostly obtained through CGMM (complex Gaussian mixture model, complex Gaussian mixture model), and then MVDR (minimum variance distortionless response, minimum variance distortionless response) is used for speech enhancement. [0003] However, the defects of the above method mainly lie in the following two aspects. The first aspect is that after the parameters of the CGMM model are randomly initialized, in order to make the CGMM model achieve better results, it is usually necessary to use the EM algorithm to iteratively update the parameters more than 20 times, so the calculation of the algorithm The complexity is very large. The second defect is that since the algorithm is performed in the frequency domain, the calculations between the fr...

AI Technical Summary

Problems solved by technology

[0003] However, the defects of the above method mainly lie in the following two aspects. The first aspect is that after the CGMM model parameters are randomly initialized, in order to make the CGMM model achieve better results, it is usually necessary to use the EM algorithm to iteratively update the parameters more than 20 times, so the calculation of the algorithm very complex

The second defect is that since the algorithm is performed in the frequency domain, the calculations between the frequency bands are independent of each other

[0006] In the implementation process of the method in the prior art, in order to ensure the later application of audio, such as recognition and other operations, it is necessary to iterate the original collected audio multiple times, so the computational complexity of the algorithm is very large

The category corresponding to each masking value is uncertain, resulting in ambiguous sorting problems

Images