| Speech is the main way of communication between people,but there are all kinds of noises in the living environment,which affect the clarity and intelligibility of speech signal and reduce the quality of speech.In order to ensure the accuracy of speech signal transmission,these noises must be suppressed.Noise is divided into steady-state noise and unsteady noise,which is divided into transient noise and other noise.Steady-state noise suppression is more mature than transient noise,which is difficult to suppress or eliminate due to its randomness and nonlinear characteristics,and is also the research difficulty and technical bottleneck of noise-containing signal.For example,in the process of human-computer voice interaction,the sudden sound of table knocking,glass breaking,keyboard sound and so on will cause great interference to the voice quality.Therefore,removing the sudden transient noise in the voice signal is an important prerequisite to improve the quality of human-computer interaction.To this kind of noise suppression effect is not ideal at present stage,the main reason is that such a sudden strong noise,energy is very big,and often and voice signal aliasing in a frequency band,the existing algorithm is hard to get to separate it from the speech signal,so the voice signal under low SNR scenario in the transient state noise eliminating research is very urgent.Firstly,in order to solve the problem of slow and under-estimation of transient noise tracking,based on the traditional Optimally Modified Log Spectral estimation(OM-LSA)method,In this thesis,the improved Mean Recurrence Time(IMRT)algorithm was used instead of the Minima Controlled Recursive Averaging(MCRA)algorithm to estimate the transient noise spectrum.Then,the optimal gain function is derived from the prior SNR and speech existence probability,so as to obtain the log-amplitude spectrum of pure speech,and then the estimated pure speech signal is obtained through inverse Fourier transform.Based on self-sampling data sets,experiments were conducted to verify the transient noise removal effect of the proposed algorithm.The results show that the proposed algorithm has Perceptual Evaluation of Speech Quality and seg SNR at-5d B.PESQ increased by 11.76 and0.97,respectively.Secondly,the existing speech enhancement methods based on deep learning send the whole speech with noise into the neural network for speech enhancement.Since transient noise does not exist in the whole time period of speech,this method is not suitable for the suppression of transient noise.In order to ensure the integrity of the speech signal,this article does not put all voice into to the neural network,but with OM-LSA algorithm to estimate the transient noise firstly,and the amplitude threshold is used to determine the existence of time domain transient noise mask,and then through the forward search as a continuous finally remove the product of speech with noise and mask time domain transient noise,It is sent to Deep Complex U-net(DCU-NET)for noise suppression to obtain the enhanced speech segment.Finally,the enhanced speech segment is inserted into the original speech sequence to obtain the enhanced speech signal.Based on the self-collected dataset,the experiment results show that the Mean Opinion Score(MOS)is all excellent,and seg SNR and PESQ can improve 13.9 and 1.46 respectively in the-5d B environment.In conclusion,the two methods proposed in this paper have good suppression effect on transient noise in speech,and significantly reduce the temporal amplitude value of transient noise.Experiments show that the proposed algorithm is effective in reducing transient noise in speech. |
PDF Full Text Request