Reading guide: Echo cancellation is the most complex module in audio call, which plays a vital role in audio call quality. all the time, Netease Yunxin It is committed to providing good echo cancellation performance and providing users with excellent audio call quality. This article will review the history of echo cancellation, analyze the current situation of echo cancellation, and look forward to the direction of further development of echo cancellation.
1、 History:
Echo is an acoustic phenomenon produced by physical and physiological interaction. After the original sound is reflected, the time delay between the reflected sound and the original sound exceeds 1 00 At ms, the human ear can hear two separate sounds, and it is easy to hear echoes in open valleys and other places.
There are generally two kinds of echoes heard during a call: line echo and acoustic echo. The line echoes because In the long-distance telephone line network, there is a problem of two to four line conversion. The local telephone is usually connected to the central office by a two-wire circuit, while the long-distance line requires a four wire circuit, and a hybrid converter is used to connect the two. The introduction of hybrid converter makes the transmitted signal generate circuit echo when the impedance of hybrid converter does not match.
The echo we mentioned in digital calls now generally refers to acoustic echo The reason for echo is that the sound from the speaker is captured by the microphone and transmitted back to the remote end, so that the remote speaker can hear his own voice.
The acoustic echo problem was not prominent in early telephone applications, mainly because of the use of low latency analog technology and the generally short communication distance. Mid 20th century in the future , the need of satellite system development raises the problem of echo cancellation, once Resolution taken one kind The method is to use the switching method to isolate the echo signal. Although this method can effectively isolate the echo, it cannot talk naturally as usual.
With the call Echoes Increasingly valued People have taken various measures to suppress or eliminate echoes. Before the 1960s, due to the limitations of objective conditions and technology, people used a method called echo suppressor to achieve echo cancellation. Echo suppression is to attenuate the echo to a certain extent to improve the communication quality, but the echo suppressor only has a good effect on a very small time delay. With satellite communications and I With the development of P telephone, there are circuits with transmission delay of more than 100-300ms. For such a large delay, the working effect of echo suppressor is very poor, so people begin to study new echo cancellation technology, and echo cancellation technology based on adaptive filter gradually becomes the mainstream direction. Since the 1980s, with the development of echo cancellation technology from simple echo suppressor to adaptive echo canceller, ITU IT U has formulated G165 (Echo Canceller) to eliminate line echo ， G167 (Acoustic Echo Canceller) for eliminating the echo of audio terminals and G 168 (Digital Network Echo Canceller) for eliminating the echo of digital networks.
present situation:
Theoretically, adaptive filter architecture is not the only choice for echo cancellation. However, due to technical limitations and historical heritage, current echo cancellation is basically based on adaptive filter architecture.
From the working principle of echo cancellation, it is a typical system identification application, identifying the echo path The convolution of reference signal and echo path is used to generate echo estimation, and then the echo estimation is removed from the mixed input 。 It is a natural choice to use adaptive filter for system identification. After a long period of technical development, researchers have made a lot of achievements in the performance research of adaptive filters At echo cancellation Common Adaptive algorithm of Deep understanding of features 。
Some commonly used Adaptive filter algorithm The characteristics are as follows:
1. LMS (least mean square) algorithm is a member of random gradient algorithm family. The algorithm uses deterministic gradient in the recursive calculation of random input Wiener filter. Its basic principle is to minimize the mean square error of the error signal core The idea is to replace mean square error with square error. The LMS algorithm is simple and computationally small, so it is easy to implement on a general digital signal processor. However, the convergence speed of the LMS algorithm depends on the eigenvalue diffusion of the autocorrelation mean of the input signal vector. When the diffusion is large, the convergence speed will be very slow. For speech signals with strong correlation, the convergence speed of LMS algorithm will be very slow 。
2. In order to overcome the shortcomings of slow convergence and sensitivity to noise of the classical LMS algorithm, the filter coefficient vector is normalized with the power of the input signal to obtain the normalized least mean square (NLMS) algorithm Law. The normalization technique reduces the dependence of the convergence rate on the input signal power, at the cost of increased computation. However, when the order is increased, the increased calculation amount is relatively small. In short, in terms of computational complexity, NLMS algorithm has the same computational efficiency as LMS algorithm. At the same time, when the input signal is speech, NLMS algorithm has faster convergence speed and better stability than LMS algorithm. Therefore, NLMS algorithm is more suitable for the application of echo canceller than LMS algorithm.
3. PNLMS algorithm and its improved algorithm are modifications to the step size control of NLMS algorithm. Its convergence speed is fast and it is effective for sparse echo paths. IPNLMS algorithm is an improvement of PNLMS algorithm, which does not require high echo path sparsity.
4. The AP algorithm can be seen as an extension of the NLMS algorithm. Because more input vectors are used, the convergence rate is faster than that of NLMS algorithm. The misalignment of AP algorithm is independent of the number of input vectors (i.e. the order of AP algorithm). With the increase of order, the convergence speed increases, but the convergence speed Degree addition The speed is decreasing. Increasing the order to improve the convergence speed is at the cost of increasing the computational complexity of the algorithm. The computational complexity of AP algorithm is much higher than that of NLMS. The convergence speed and computational complexity of AP algorithm are between NLMS algorithm and RLS algorithm.
5. Unlike random gradient algorithm, RLS algorithm belongs to least square algorithm. The convergence speed of RLS algorithm is one order of magnitude faster than that of LMS algorithm. As the number of iterations approaches infinity, the additional mean square error of RLS algorithm converges to zero. However, because of the use of matrix inversion, the RLS algorithm has a large amount of computation, which is difficult to achieve for the elimination of echo signals whose impulse response exceeds 100 milliseconds.
Most of the current echo cancellation modules are based on one of the above algorithms or their variants, which is close to the ideal situation , existing adaptive filtering algorithm all Can get relatively Satisfactory echo cancellation effect However, this is only an ideal state in most of the actual calls, and the actual calls are much more complicated. Local noise interference, nonlinear transformation in signal path, simultaneous speech of both parties, data delay jitter caused by non real-time operating system, etc. are all problems to be solved by general echo cancellation module.
for example When there is obvious near end sound, the performance of echo cancellers based on existing adaptive filtering algorithms will deteriorate, and even the convergence of adaptive filtering algorithms cannot be guaranteed. This is the key problem that must be solved in the practical application of echo cancellation, usually called double talk Question. Currently, the most common way to solve this problem is to add a Double Talk Detector (DTD) , when both sides speak at the same time, let the adaptive filter lock or slow change To prevent the adaptive process from diverging.
How to give satisfactory logic and fine adjustment to these additional problems in the case of real-time calls is very important for productization, and it is also the main reason for product performance differences. The echo cancellation module in the current commercial products can achieve satisfactory results in most cases, but in some special cases there will be some problems. The requirement for higher call quality is still the driving force to research better echo cancellation modules.
future:
Due to the complexity of the call echo problem, there is always a need to improve the effect of the echo cancellation module. The inherent problems of the current scheme make it difficult to make breakthrough progress on the above-mentioned key issues. Due to the improvement of equipment computing power and the development of technology in recent years, some new schemes have also been proposed. With Vol Terra filter Based on Of Nonlinear filter solutions, research on echo cancellation algorithms based on ICA and deep learning are also emerging, and echo cancellation algorithms based on linear adaptive filter architecture are facing challenges. However, no other architecture can be widely used like the current adaptive filter architecture, and echo cancellation algorithm will still exist as a common technical topic in the short term.
Netease Yunxin continues to promote research on the improvement of echo cancellation logic and new scheme research, and is committed to providing users with better audio call effects.
understand Netease Yunxin Communication and video cloud services from NetEase core architecture>>
More technical goods, please pay attention to vx official account "Netease Smart Enterprise Technology+" 。 A series of courses can be viewed in advance. Boutique gifts are free, and you can also talk to CTO directly.
