Telephone Conferencing Systems, Multi-channel Echo Cancellation Theory And Applied Research

In teleconferencing systems, the echo is inevitable. With the wide application of teleconference and hand-free systems, the acoustic echo cancellation becomes a heat topic at home and abroad. However, the research achievements of multi-channel acoustic echo cancellation are much less compared to those of mono-channel. There are many theories, approaches and implementation techniques to be studied.When stereo signals from the same source is transmitted, the solution to the normal equation of the adapt filter is singular, so echo cancellation can not provide the unique solution to the echo path. In addition, the key problem of multi-channel echo cancellation lies in the strong correlation among input signals. Due to the strong correlation, the condition of the correlation matrix gets inferior, and the divergence of eigenvalue becomes large. This will reduce the convergence rate in general gradient adaptive algorithm, which makes the stability of all kinds of adaptive recursive algorithm inferior. Correlation among input signals also causes imbalance in the coefficient vector.The convergence speed and computational complexity about adaptive algorithms is the core issue about echo cancellation in teleconferencing systems. If the convergence speed is very slow or computational complexity is very high, adaptive algorithms will not track the changes in the echo channel in time, and the residual echo will badly affect the conference effect. Usually, convergence speed increases at the expense of computational complexity. To ensure the convergence rate and reduce the computational complexity, this paper will find a balance between the convergence speed and computational complexity as far as possible to arrive at best Real-time effect.To break the traditional thought way, we don't use the same approach to deal with the same voice signal in this essay, but make full use of the distribution characteristic of voice signals in the frequency, that is, low-frequency with more energy, high-frequency with less energy. As an innovation, the different adaptive approaches are taken in different frequency sections of voice signal in this paper.After comprehensive analysis of the two-channel affine projection algorithm and normalized least mean square error algorithm, we find that the two-channel affine projection algorithm has a faster convergence speed but more complicated than normalized least mean square error algorithm. Besides, analysis of the filter bank (sub-band filter) in different channels were taken and the signal adaptive filter is got, so that each sub-band adaptive filter has a shorter impulse response than full-band filter, and computational complexity has been significantly reduced. Based on this analysis, taking advantage of the combination of the two-channel affine projection algorithm, two-channel normalized least mean square error algorithm, the analysis filter group and the distribution characteristic of voice signals in the frequency, this article comes up with Combination algorithm and conduct an experiment to validate the results. The results show that the combination method has the similar convergence rate with affine projection, and the computational complexity is obviously lower, so the combination method is perfectly real-time in echo cancellation system.