| In the field of acoustic signal processing,the technique of using microphone arrays to locate sound sources is a research hotspot and is widely used in video conferencing,hands-free voice communication,human-computer interaction,and voice enhancement.In a real room acoustic environment,the mainly processed sound is usually human speech,which has its unique properties.The background noise and reverberation in the room will degrade the quality of the speech signals received by the microphones,which could increase the difficulty of speech source location.Therefore,reliable speech source localization in complex room acoustic environments remains challenging.Linear prediction is an important pre-whitening technique,which whitens the background noise to enhance performance of TDE.Common linear prediction methods construct a least squares model to solve an optimization issue.But this model is not optimal for speech signals due to their non-Gaussian distribution.The model that use the-norm of the predictor coefficient vector and the-norm of the prediction residual vector could effectively improve the performance.In this paper,the linear prediction algorithms are used for time delay estimation(TDE)to carry out sound source localization,and the three main research works are carried out as follow:1)In the process of solving the sparse model using the-norm,it is inevitable to invert the high-dimensional cross-correlation matrix,which results huge calculations.Aiming at the shorting of this model,a TDE algorithm with low computational complexity is proposed in this paper.Considering the sparsity of speech signals,the linear predictor is decomposed into two groups of short subpredictor by using the Kronecker product,which converts the high-dimensional problems into low-dimensional ones,and the computational load of the algorithm is thereby greatly reduced.The simulation experiment verified that the proposed algorithm can maintain the TDE performance while reducing the calculations.2)The amplitude of the short-time Fourier transform(STFT)of the pure speech signal is sparse,and experience shows that the amplitude spectrum of the STFT of the microphone signal is also sparse.The short-term change trend of the microphone signal is often predictable.Therefore,this paper proposes a time delay estimation algorithm with speech spectrum constraints.Taking advantage of the sparsity of the speech spectrum,the convex constrained linear prediction model is defined by using the-norm of whitened signal vector and the-norm of the prediction error vector,which can effectively compromise the pre-whitening of the microphone signal,and gain robustness to noise and reverberation.Simulation experiments verified the effectiveness of the proposed algorithm under different conditions.And the advantages are more significant in the environment dominated by reverberation.3)Since the time delay estimation algorithm constrained by the speech spectrum effectively utilizes the frequency domain characteristics of the speech signal,it achieves relatively better performance,but its solution process requires Fourier transform and inverse Fourier transform as well as high-dimensional matrix inversion,which leads to a huge amount of calculation.So the Kronecker product decomposition is introduced so that the linear predictor is decomposed into two groups of short sub-predictor by using the Kronecker product,which converts the high-dimensional problems into low-dimensional ones,and the computational load of the algorithm is thereby greatly reduced.to reduce the amount of calculation,so that the advantages of both the performance of the algorithm and the speed of operation can be taken into account.Simulation experiments verified the effectiveness of the proposed algorithm. |
