Study On Adaptive Echo Cancellation And Noise Reduction

IntroductionIn the communication systems, echo and noise will seriously affect speech communication quality, such as IP telephone, videophone and hand-free digital cell phone systems etc. Therefore, echo cancellation and background noise reduction is one of the key technologies for improving speech quality. With the development of echo cancellation technologies, the research emphases for echo cancellation has already been turned from network echo to acoustics echo at present. The studies of adaptive filters begin at the end of the fifties in last century. Windrow and Hoff etc. first put forward adaptive algorithms of least minimum square (LMS). Since LMS algorithm is simple, small in calculation complexity, easy to realize in real time signal processing, it is widely used in many fields. When it is used in echo cancellation, for non-stationary and strong relevant speech signal, and echo paths with long impulse responses, LMS algorithm directly realizes the estimation of echo path in time domain will result in over-stress burden of calculation and low convergence rate. Various kinds of improvement algorithm for LMS have been put forward, and these algorithms can be regarded as the methods of step-size parameter varying. It has already proved that the value of step size parameter is in inverse proportion to time constant of studying curve and in direct proportion to misadjustment, namely optimization of step-size parameter is the trade off between convergence speed and small misadjustment. This kind of thought is the basis of improvement of various kinds of algorithms such as NLMS. NLMS overcomes the coherence problem of adapting and controlling brought from large step size parameter and input signal power in LMS and still keeps simpleness of the original LMS algorithm. Since later seventies in the 20th century, NLMS has become the algorithm that commercialized echo cancellers are often adopted. But its shortcoming is that the convergence performance of the error signal would worsen sharply under the condition that the strong correlative speech signal inputs.In order to satisfy the requirement of echo canceller, NLMS must be improved. Theoretically echo canceller needs long echo path of impulse response, but in fact only few non-zero areas in the whole path. PNLMS algorithm is a kind of new adaptive scheme. It converges more quickly than NLMS through making use of sparseness of echo path. It improves the algorithm performance while it does not reduce the quality of echo estimation. But the premise of using the algorithm is that echo path is sparse. The modified algorithms from NLMS,PNLMS such as RPNLMS, PANLMS, RPNLMS and NRPNLMS promote slightly the performance with little increasing of calculating complexity. Affine projection algorithm (APA) processes higher parameter accuracy and convergence speed for echo path estimation. It can get better performance than those algorithms through properly selecting projection order. At present, most echo cancellers use NLMS algorithm to realize echo cancellation, and do not conduct cancellation of noise that exists in echo path. Hence it will exert effect on echo cancellation. Considering the purpose speech communication, it is necessary to cancel echo and reduce noise at the same time.This paper is based on above discussed thoughts. In front of echo cancellation, the input speech signal is enhanced, which establishes the foundation for the following-up echo cancellation. Then good performance of echo cancellation could be gotten after the enhanced speech entering echo canceller with DTD. Whole algorithm becomes more integrated and robust through the combination structure of noise reduction and echo cancellation.1. Noise reduction and echo cancellationVoice activity detector in spectral subtractionSpectral subtraction is an effective method of noise reduction in frequently domain. This algorithm is not only simple, fast, low operation complexity, and easy to realize but also can reach higher output SNR. In fact, a lot of telephone products use spectral subt...