Author: Wang Jiajie
近日，韩国汉阳大学电子工程系JungPhil Park等人提出一种新型ANC架构，融合自适应滤波（Adaptive Filter, AF）与深度神经网络（Deep Neural Networks, DNN），简称HAD-ANC，以解决前馈ANC问题。 具体而言，AF部分选用归一化频域块最小均方差（Normalized Frequency Block Least Mean Square, NFBLMS）算法，DNN部分选用两个门控卷积循环神经网络（Gated Convolutional Recurrent Network, GCRN）模块。
In the feedforward ANC algorithm, as in the above formula, E is the error signal picked up by the physical error microphone, which is the sum of the initial noise signal and the secondary noise signal. X is the feedforward signal picked up by the reference microphone, P is the primary channel, C is the feedforward controller, and S is the secondary channel.
The optimal solution of the controller is the inverse of the primary channel P multiplied by the secondary channel S. The secondary channel includes not only the physical secondary channel, but also the electroacoustic (horn, speaker, etc.) channel, which is expressed separately from the physical secondary channel in the paper, but the essence is the same.
In the proposed scheme, the AF part is responsible for modeling the primary channel P, and the GCRN1 module in the NN part models the inverse of the secondary channel S. (Therefore, AF and GCRN1 are merged as feedforward controller C.) GCRN2 models the secondary channel S to estimate the primary channel with a forced AF part. Therefore, in this scheme, the error signal is:
If the feedforward ANC only contains AF, it is not only adaptive, but also simple and efficient, but it is essentially a linear model, and the nonlinear part cannot be modeled, and the amount of noise reduction is limited. If feedforward ANC only contains NN, although it has strong nonlinear modeling capabilities, it consumes too many parameters and calculations for the modeling of the linear part, and the cost performance is low.
The cascading scheme architecture proposed in this paper can complement the advantages of AF and DNN: AF can model the primary channel linearly, with a small number of parameters and calculations, cost-effective, and has the ability to learn and converge after channel mutation, and the network part can model the secondary channel (including the nonlinear part of the speaker) with strong nonlinear ability to obtain a high-performance controller.
Figure 1 Comparison of noise reduction between the proposed algorithm and other algorithms before and after channel mutation (Fig. 9)
As shown in the figure above, the noise reduction performance of the proposed algorithm is comparable to that of Deep-ANC [2] proposed by Hao Zhang before the channel mutation. Because HAD-ANC contains AF, it takes a certain initial time to converge and achieve the same noise reduction performance as Deep-ANC later, and other traditional algorithms and SPD-ANC are inferior to Deep-ANC and HAD-ANC.
However, after the channel mutation, Deep-ANC does not contain an adaptive module, and if the randomization generated in the training data is too different from the measured mutated signal, it cannot be effectively extrapolated, resulting in a direct collapse of performance, and a rebound instead of noise reduction. On the other hand, HAD-ANC, with its AF part, has the ability to adapt to channel mutations, and it is still better than other algorithms after convergence.
It should be noted that the reason why the SPD-ANC algorithm also has the ability to reconverge after channel mutation is that the algorithm also includes the AF module, as detailed in the paper by Daocheng Chen of the Institute of Acoustics of the Chinese Academy of Sciences [3].
An algorithm similar to the ANC principle is AEC. In the application scenario of mobile phone VOIP or front-end cabin, the echo path may also change abruptly due to the movement of people or the opening and closing of windows and doors. In the industrial world, the traditional AF+NN technology solution for treating residual echo has long been popular and occupies the mainstream.
bibliography
[1] Park J P, Choi J H, Kim Y, et al. HAD-ANC: A Hybrid System Comprising an Adaptive Filter and Deep Neural Networks for Active Noise Control[C]//Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH. International Speech Communication Association, 2023, 2023: 2513-2517.
[2] Zhang H, Wang D L. Deep ANC: A deep learning approach to active noise control[J]. Neural Networks, 2021, 141: 1-10.
[3] Chen D, Cheng L, Yao D, et al. A secondary path-decoupled active noise control algorithm based on deep learning[J]. IEEE Signal Processing Letters, 2021, 29: 234-238.