Research On Algorithms Of Speech Enhancement In The Low SNR And Complicated Environments

With the development of science and technology,people have put forward higher requirements for the quality of speech communication.The process of speech communication will be interfered by various noises at the same time.In a low signal-to-noise ratio(SNR)environment,the speech will be submerged by noise,which will seriously affect the quality of speech communication,and even make it impossible to communicate.Therefore,highly robust speech enhancement technology is becoming more and more important.The multi-channel speech enhancement algorithm introduces spatial information,realizes filtering in the spatial domain,and improves the speech enhancement performance.However,this algorithm has strict requirements on the size,number,and location of the microphone array,and requires high equipment maintenance,and does not use the application of low-cost equipment.Therefore,this thesis studies the single-channel speech enhancement algorithm,focusing on the nonnegative matrix factorization(NMF)algorithm based on sparse constraints,and proposes improvements to the algorithm.For the improved algorithm,the project implements a real-time online speech enhancement system.Traditional speech enhancement algorithms need to use noise estimation algorithms to estimate the noise,but the signal-to-noise ratio and noise estimation errors in a complex noise environment are large,resulting in a reduction in the speech enhancement performance of traditional algorithms.The NMF algorithm does not require noise estimation.According to the noise dictionary and speech dictionary trained before enhancement,the noise component and speech component are solved by matrix decomposition of the noise and speech power spectrum amplitude,so as to enhance the speech,and the speech enhancement performance is better than the traditional algorithm.This thesis focuses on the improved algorithm of the NMF algorithm-NMF algorithm with sparseness constraints,discusses whether the noise matches the performance of NMF algorithm with sparseness constraints,and conducts experimental verification.Combining the characteristics of Mel spectrum and the principle of non-negative matrix factorization,different from the traditional matrix factorization using ordinary amplitude spectrum as data,a method of matrix factorization based on Mel spectrum as data is adopted,and a NMF algorithm with sparseness constraints based on Mel spectrum is proposed.Experiments have proved that the PESQ score of the improved algorithm is better than that of the traditional algorithm,and the algorithm complexity is greatly reduced.The existing NMF algorithm with sparseness constraints uses a fixed noise dictionary and a speech dictionary.When the noise of the noisy speech and the noise dictionary do not match,the speech enhancement performance decreases.Based on this problem,this thesis improves the original algorithm,designs and implements a speech enhancement system based on the NMF algorithm with sparseness constraints.The improvement is divided into two aspects.On the one hand,the filter module is designed and the sparseness measure method is introduced.Calculate the probability of speech existence through sparseness,use speech existence probability and posterior signal-to-noise ratio to design an adaptive parameter change factor to improve the original filter.The designed parameter change factor can be adaptively adjusted in different frequency bands and signal-to-noise ratio.Experiments show that The PESQ score of the improved filter's NMF algorithm with sparseness constraints is better than that of the traditional algorithm.On the other hand,an adaptive noise update module is designed to selectively update the noise dictionary using the calculated speech existence probability.Experiments have proved that the PESQ score of the adaptive noise update module under low signal-to-noise ratio is improved to a certain extent,and the noise dictionary is enhanced.Robustness.Finally,on the Raspberry Pi platform,a speech enhancement system based on NMF algorithm with sparseness constraints was engineered.The designed speech enhancement system is divided into two large modules,a training module and an online enhancement module.The training module implements batch training audio.The online enhancement module realizes batch processing enhanced audio and online enhanced audio.The online enhanced audio adds a recording reading module and a recording processing module.The openblas matrix operation library optimizes matrix operations,optimizes memory structure,optimizes code structure,etc.to achieve algorithm optimization,Which reduces the calculation time of the algorithm implemented in the project,and realizes the real-time online enhanced audio processing function,and tested by the Raspberry Pi,which proves that the calculation time of the algorithm is greatly reduced after the optimization.