Real-time Noise Reduction For Conference Telephone

Conference telephone is an essential tool in business communication.However,noise in the Conference telephone can lead to a serious deterioration in the quality of communication among the participants.Therefore,noise reduction is of great significance to the design and development of conference telephone.The real-time noise reduction algorithm in conference telephone system and real-time noise reduction algorithm based on deep learning are both studied.The theoretical knowledge,specific research methods and test results of noise reduction algorithm are proposed.The main work is as follows:1)The optimally-modified log-spectral amplitude(OM-LSA)speech estimator is studied.Firstly,the traditional log-MMSE estimator and the OM-LSA estimator are introduced.Then,different noise estimation algorithms are studied and compared experimentally to confirm that IMCRA algorithm is the optimal noise estimation algorithm.Finally,by comparing different estimators,it is confirmed that the OM-LSA estimator is the best one.2)A real-time noise reduction algorithm based on multi-characteristic modified estimator is proposed.Firstly,a method of estimating stationary noise by using various types of noise is proposed.Then,the transient noise is estimated by using the algorithm of estimating the transient noise.Meanwhile,a method of estimating the probability of speech existence by using various speech features and harmonics is proposed.Finally,the OM-LSA estimator is used to suppress the noise.In the experimental simulation stage,it is confirmed that the existence probability method based on harmonic can improve speech quality,and the performance comparison of several noise reduction algorithms confirms that the algorithm has good suppression effect on both stationary and non-stationary noise.3)A real-time noise reduction algorithm based on RNN is proposed.Firstly,the Bark spectral coefficient(BFCC)is obtained by dividing the frequency band based on Bark domain,which is used as input feature to reduce the input of training GRU model.Then,the gain is obtained by the model and speech enhancement is achieved by comb filter filtering.In the experimental simulation stage,it is confirmed that the RNN-based noise reduction algorithm has a good denoising effect.Finally,the real-time test shows that the proposed algorithms can meet the real-time requirements of conference telephone.