自适应波束形成（四）——Frost波束形成1

最新推荐文章于 2024-07-10 17:28:54 发布

叶夜笙歌

最新推荐文章于 2024-07-10 17:28:54 发布

阅读量1.2w

点赞数 4

分类专栏：语音增强

本文链接：https://blog.csdn.net/YJJat1989/article/details/22398831

版权

语音增强专栏收录该内容

9 篇文章 34 订阅

订阅专栏

1 算法原理

LCMV波束形成器最优权向量的求解依赖于阵列接收数据的统计知识，即要计算自相关R_x。然而，实际应用中阵列接收数据的自相关是未知的或是时变的，此时约束自适应算法能用来逼近权向量，即Frost波束形成器。

首先，初始化权向量w[0] = C(C^HC)^-1f，它满足约束条件。然后以一个适当的比例系数朝R_xxw+Cλ 负梯度方向迭代，得到

w[n+1] = w[n] - µ(R_xxw[n] +Cλ[n])

既然w[n+1]满足C^Hw = f ，将其代入解得λ[n]，然后代回上式，得到

w[n+1] = C(C^HC)^-1f +P(w[n]-µR_xxw[n])

其中P=I- C(C^HC)^-1C^H，I为单位矩阵。

由于不知二阶统计量 R _xx，自相关矩阵可近似为 R _xx≅ xx ^H。而 x ^H w[ n]恰是阵列第n次迭代的输出y[ n]，因此Ftost波束形成器归结为

w[0] = C(C^HC)^-1f

w[n+1] = C(C^HC)^-1f +P(w[n]-µy[n]x[n])

2 对模拟麦克风阵列语音进行增强

在matlab r2011a的toolbox的phased有波束形成相关的demo，demo之一代码如下：

%Define a Uniform Linear Array 
%First we define a uniform linear array (ULA) to receive the signal. 
%The array contains 10 omnidirectional microphones and the element spacing is 5 cm
hmic=phased.OmnidirectionalMicrophoneElement;
ha=phased.ULA(4,0.05,'Element',hmic);
c=340;                                       %sound speed,in m/s

%Simulate the Received Signals
load('twospeeches','speech1','speech2');
load('laughter','y');                       %The laughter is stored in variable y
y=2*y*(1:length(speech1));                   %Amplify and truncate
fs=8192;                                    %in Hz

wavwrite(speech1,fs,'speech1.wav');
wavwrite(speech2,fs,'speech2.wav');
ang1=[-30;0];
ang2=[60;10];
angInt=[20;0];

hCollector=phased.WidebandCollector('Sensor',ha,'PropagationSpeed',c,...
     'SampleRate',fs,'ModulateInput',false);
 sigSource=step(hCollector,[speech1 speech2 y],[ang1 ang2 angInt]);
 
 rs=RandStream.create('mt19937ar','Seed',2008);
 noisePwr=1e-4;                              %noise power
 sigNoise=sqrt(noisePwr)*randn(rs,size(sigSource));
 
 sigArray=sigSource+sigNoise;
 wavwrite(sigArray(:,3),fs,'sigArray.wav');
 
 %Plot channel 3
 plot(sigArray(:,3));
 xlabel('Time(sec)');ylabel('Amplitude(V)');
 title('Signal Received at Channal 3');ylim([-3 3]);
 
 %Listen to channal 3
 player=audioplayer(sigArray(:,3),fs);
 play(player);
 
 %Process with a Time Delay Beamformer
 angSteer=ang1;
 hbf=phased.TimeDelayBeamformer('SensorArray',ha,'SampleRate',fs,...
     'Direction',angSteer,'PropagationSpeed',c)
 
 cbfOut=step(hbf,sigArray);
 
 plot(cbfOut);
 xlabel('Time(Sec)');ylabel('Amplitude(V)');
 title('Time Delay Beamformer Output');ylim([-3 3]);
 
 player=audioplay(cbfOut,fs);
 play(player);
 
 agCbf=pow2db(mean((speech2+y).^2+noisePwr)/mean((cbfOut-speech1).^2))
 
 %Process with a Frost Beamdormer
 hbf=phased.FrostBeamformer('SensorArray',ha,'SampleRate',fs,...
     'PropagationSpeed',c,'WeightsOutputPort',true);
 %The beamformer may change its steering direction during processing
 hbf.DirectionSource='Input port';
 hbf.FilterLength=20;    %Set the length of FIR filter for each sensor to 20
 
 [FrostOut,w]=step(hbf,sigArray,ang1);
 
 plot(FrostOut);
 xlabel('Time(sec)');ylabel('Amplitude(V)');
 title('Frost Beamformer Output');ylim([-3 3]);
 
 player=audioplayer(FrostOut,fs);
 play(player);
 
 agFrost=pow2db(mean((speech2+y).^2+noisePwr)/mean((FrostOut-speech1).^2))
 
 %Use Diagonal Loading to Improve Robustness to the Frost Beamformer
 release(hbf);
 dir2=[50;20];               %Estimated steering direction
 FrostOut=step(hbf,sigArray,dir2);
 
 plot(FrostOut);
 xlabel('Time(sec)');ylabel('Amplitude(V)');
 title('Frost Beamformer Output');ylim([-3 3]);
 
 player=audioplayer(FrostOut,fs);
 play(player);
 
 agFrost2=pow2db(mean((speech1+y).^2+noisePwr)/mean((FrostOut-speech2).^2))
 
 release(hbf);
 hbf.DiagonalLoadingFactor=1e-3;    %Specify diagonal loading value
 FrostOut=step(hbf,sigArray,dir2);
 
 plot(FrostOut);
 xlabel('Time(sec)');ylabel('Amplitude(V)');
 title('Frost Beamformer Output');ylim([-3 3]);
 
 player=audioplayer(FrostOut,fs);
 play(player);
 
 agFrostDL=pow2db(mean((speech1+y).^2+noisePwr)/mean((FrostOut-speech2).^2))

首先定义一个均匀线性阵列，包含8个全向性麦克风阵元，阵元间距5cm，声速340m/s；然后模拟麦克风接收到的信号，语音1为一段男声，语音2为一段女声，采样率8192Hz，分别在-30^o和60^o，同时在20^o有一段笑声作为干扰，并加上随机噪声作为本底噪声，用matlab中函数模拟产生一个39922*8的矩阵，每一列代表一个阵元接收到的混合声音信号。各信号如图1所示。