Research Of Speech Enhancement Technology Based On Intelligent Appliance

In recent years,we have made a breakthrough in smart home.Mobile voice recognition and translation system play an important role in smart home,but complex and changeable environmental noise is one of the main obstacles to the effect of speech recognition technology.The purpose of speech enhancement is to improve the quality and intelligibility of these damaged speech.Minimum mean square error(MMSE)of speech enhancement algorithm has relatively effective denoising performance for stationary and non-stationary noise,but it always distorts voice seriously.Traditional signal subspace algorithm can divide the speech signals which with the noise into a pure voice subspace and noise subspace,and almost no damage in the process of denoising.This algorithm is effective to denoise stationary noise obviously,but the algorithm can hardly estimate the dimensions of the subspace,and has poor denoising performance to non-stationary noise.To solve this problem,this thesis proposes an improved dual-channel signal subspace speech enhancement algorithm based on the advantages of MMSE algorithm and signal subspace algorithm.This thesis uses the adaptive MMSE algorithm to estimate time delay,this algorithm can adjust the filter coefficients according to the sampling of the current input signal to minimize the output signal error without prior knowledge of the input signal spectrum.The eigenvalue decomposition(EVD)of the noisy speech signal combined with the dual channel processing method process the input speech signal in the beam-forming way,without estimation of the signal subspace dimension.In order to further suppress the coherent noise,the minimum mean square error constraint is introduced in the dualchannel subspace.This method using the largest eigenvalue in the noise signal subspace as the estimate of the noise variance,can not only maximize the effect for denoising,but also do little harm to the speech when estimating the useful speech signal.The simulation results show that the algorithm can not only remove the white noise effectively,but also has a good noise reduction effect on the colored noise.The noise used in this thesis is selected from the NOISE92 noise library which closer to the noise type of the home environment,the other type of noise is recorded in the actual home environment.Finally,this algorithm is applied to the offline speech recognition system in the home environment,which proves that the algorithm can improve the recognition performance of the speech recognition system in a variety of noise environments.