Speech Enhancement Method Based On Subband Kalman Filter Combined With Phase Reconstruction

Speech enhancement is the removal of noise from noisy speech to improve speech quality and intelligibility.There are several speech enhancement algorithms including Kalman filter,among which the enhancement performance of Kalman filter depends largely on the estimation accuracy of the parameters,but there are shortcomings in the estimation of parameters in the traditional Kalman filter method,which directly affects the enhancement performance of Kalman filter.In addition,most speech enhancement methods only process the amplitude spectrum when enhancing noisy speech and the phase is directly replaced by the phase of noisy speech,because earlier researchers believed that the contribution of phase to speech quality is limited.Recent studies have shown that phase has a certain effect on speech quality improvement and the reasonable use of phase can further improve the quality of enhanced speech.In order to solve the problems arising from the above-mentioned methods for speech enhancement,this paper investigates the speech enhancement method based on subband Kalman filtering combined with phase reconstruction.Specific research contents are as follows:(1)The research value of speech enhancement techniques and the status of domestic and international research are introduced.Common concepts of speech signals are described and deep learning models related to the methods of this paper are briefly described.(2)A subband Kalman filter speech enhancement method based on multi-target deep neural network is proposed to address the problem that the traditional Kalman filter cannot accurately estimate the linear prediction coefficients(LPC)under the actual noise conditions leading to the degradation of the enhancement performance when the Kalman filter is used.This method decomposes the noisy speech into noisy subband speech by discrete wavelet transform(DWT),calculates the LPC of noisy subband speech and transforms them into line spectrum frequencies(LSF),takes the LSF of noisy subband speech as the input of the network and the corresponding LSF of clean subband speech and noisy subband as the output of the network.The output LSF are converted into LPC and used to construct a subband Kalman filter and the noisy subband speech is subband Kalman filtered to obtain the enhanced subband speech and the enhanced subband speech is then synthesized into enhanced fullband speech by inverse discrete wavelet transform(IDWT).The experimental results show that the proposed method outperforms several Kalman filter-based comparison methods in terms of speech quality and intelligibility.(3)To address the shortcomings of traditional speech enhancement methods in phase processing and the prevalence of speech distortion issue in the enhancement process,a speech enhancement method with improved phase compensation combined with harmonic reconstruction is proposed.The method estimates the a priori signal-noise-ratio(SNR)by deep learning model and improve the traditional phase spectral compensation function by using the a priori SNR,on this basis,the enhanced speech is secondarily enhanced by harmonic reconstruction to solve the existing speech distortion issue.The experimental results show that the speech enhancement method with improved phase compensation combined with harmonic reconstruction has better enhancement capability than the comparison methods,which can effectively reduce speech distortion and improve speech quality.