自適應回聲消除算法

歡迎留言交流

AEC算法早期用在Voip，電話這些場景中，自從智能設備誕生後，智能語音設備也要消除自身的音源，這些音源包括音樂或者TTS機器合成聲音。

本文基於開源算法闡述AEC的原理和實現，基於WebRTC和speex兩種算法，文末會附上兩種算法的matlab實現。

回聲消除原理

回聲消除的基本原理是使用一個自適應濾波器對未知的回聲信道:ωω 進行參數辨識，根據揚聲器信號與產生的多路回聲的相關性爲基礎，建立遠端信號模型，模擬回聲路徑，通過自適應算法調整，使其衝擊響應和真實回聲路徑相逼近。然後將麥克風接收到的信號減去估計值，即可實現回聲消除功能。
echo=x∗ωecho=x∗ω 1.1
d=s+echod=s+echo 1.2
y^=x∗ω^y^=x∗ω^ 1.3
e=d−y^e=d−y^ 1.4
式中ωω是回聲通道的時域衝擊響應函數，x是遠端語音；echo是所得回聲；s是近端說話人語音，d爲麥克風採集到的信號，y^y^是對回聲信號的估計值，e爲誤差。
爲了消除較長時間的回聲，需要FIR濾波器的階數較高，時域計算法，有兩個問題，一個是實時性較差，一個是計算量大。爲了在實時性/計算量以及可以消除的回聲時長之間找到使這三個最優的算法，採用了頻譜分塊自適應濾波算法。
這裏用到了很多信號處理算法，爲了讓算法理解起來容易些，簡單羅列涉及到的算法：

FFT/IFFT
循環卷積和線性卷積的關係;重疊保留法
功率譜密度
互相關
NLMS自適應算法

NLMS權重調整

關於NLMS，可以下載http://download.csdn.net/detail/shichaog/9832657

下面直接開始WebRTC的matlab梳理，由於matlab代碼和webRTC的c++代碼命名幾乎一致。所以c++的實現就一筆帶過。
首先解釋幾個名詞：

RERL-residual_echo_return_loss
ERL-echo return loss
ERLE echo return loss enhancement
NLP non-linear processing

首先matlab讀入遠端和近端信號。

%near is micphone captured signal
fid=fopen('near.pcm', 'rb'); % Load far end
ssin=fread(fid,inf,'float32');
fclose(fid);
%far is speaker played music
fid=fopen('far.pcm', 'rb'); % Load fnear end
rrin=fread(fid,inf,'float32');
fclose(fid);

然後對一些變量賦初值

fs=16000;
NLPon=1; % NLP on
M = 16; % Number of partitions
N = 64; % Partition length
L = M*N; % Filter length
VADtd=48;
alp = 0.15; % Power estimation factor 
alc = 0.1; % Coherence estimation factor
step = 0.1875;%0.1875; % Downward step size

上述初始化中，M=16和最新的WebRTC代碼並不一致，且最新的WebRTC中支持aec3最新一代算法。

len=length(ssin);
NN=len;
Nb=floor(NN/N)-M;
for kk=1:Nb
    pos = N * (kk-1) + start;

可以看出Nb是麥克風採集到的數據塊數-16（分區數），這是因爲第一次輸入了16塊，所以這裏減掉了16。pos是每一次添加一塊時的地址指針。

    %far is speaker played music
    xk = rrin(pos:pos+N-1);
    %near is micphone captured signal
    dk = ssin(pos:pos+N-1);

xk和dk是讀取到的64個點，這裏是時域信號。

功率計算

    %----------------------- far end signal process
    xx = [xo;xk];
    xo = xk;
    tmp = fft(xx);
    XX = tmp(1:N+1);

    dd = [do;dk]; % Overlap
    do = dk;
    tmp = fft(dd); % Frequency domain
    DD = tmp(1:N+1);

將xk和上一次的數據結合在一起，做FFT變換，由於兩次組合在一起，那麼得到的應該是N=128點，這裏可以知道得到的譜分辨率是n∗fs/Nn∗fs/N,fsfs前面設置過了，是16k，則得到的每一個bin的頻譜分辨率是16000/128=125Hz。這裏XX和DD取了前65個點，這是因爲N點FFT變換後頻譜是關於N/2對稱的，對稱的原因是奈奎斯特採樣定理，如果fs=16000Hzfs=16000Hz,那麼要求採樣到的信號的截止頻率必然小於等於fs/2=8000Hzfs/2=8000Hz,對於實信號，N/2~N,實際上表示的是−fs/2−fs/2 ~ 00之間的頻率。第一個點是直流分量，所以取65個點。和上一幀64個點信號合併在一起的另一個原因是使用重疊保（overlap-save）留法將循環卷積變成線性卷積，這裏做的FFT變換，就是爲了減少時域裏做卷積的計算量的。
計算遠端信號功率譜

    % ------------------------far end Power estimation
    pn0 = (1 - alp) * pn0 + alp * real(XX.* conj(XX));
    pn = pn0;

平滑功率譜，上一次的功率譜佔85%（alp=0.15），後面的頻域共軛相乘等於功率是有帕斯瓦爾定理支撐的。pn0是65*1的矩陣。

濾波

 XFm(:,1) = XX;

首先將遠端信號頻譜賦給XFm，XFm是65*16的矩陣，16就是前面初始化的M值，這裏將XX給第一列，其2~16列對應的是之前的輸入頻譜。

    for mm=0:(M-1)
        m=mm+1;
        YFb(:,m) = XFm(:,m) .* WFb(:,m);
    end

YFb，WFb以及XFm都是65*16的矩陣，WFb是自適應濾波器的頻譜表示，XFm是原始的speaker數據，上式的意義對應於插圖中的y^y^的頻域值，變換到時域後就可以得到yy的估計值y^y^.

    yfk = sum(YFb,2);
    tmp = [yfk ; flipud(conj(yfk(2:N)))];
    ykt = real(ifft(tmp));
    ykfb = ykt(end-N+1:end);

首先yfk是65*1的矩陣，sum求和就是將估計的頻譜按行求和，也就是yfk包含了最近16個塊的遠端頻譜估計信息，這樣，只要近端麥克採集到的信號裏有這16個塊包含的遠端信號，那麼就可以消掉，從這裏也可以看出來，容許的延遲差在16*64/16=64ms，也就是說，如果麥克風採集到的speaker信號滯後speaker播放超過64ms，那麼這種情況是無法消掉的，當然，延遲差越小越好。
flipud(conj(yfk(2:N))是因爲前面計算頻譜時利用奈奎斯特定理，也即實數的FFT結果按N/2對稱，所以這裏爲了得到正確的ifft變換結果，先把頻譜不全到fsfs.
ykfb就是y^y^.後面再看WFb是如何跟新。

誤差估計

   ekfb = dk - ykfb;

dk是麥克風採集到的信號，ykfb是y^y^,這樣得到的是誤差信號，理想情況下，那麼得到的誤差信號就是需要的人聲信號，而完全濾除掉了speaker信號（遠端信號）。

    erfb(pos:pos+N-1) = ekfb;
    tmp = fft([zm;ekfb]); % FD version for cancelling part (overlap-save)
    Ek = tmp(1:N+1);

erfb是近端信號數組長度×1矩陣，存放的是全部樣本對應的誤差信號，這個保存僅僅是爲了plot用的。
然後補了64個零，然後做FFT，Ek是誤差信號FFT的結果。

自適應調節

   Ek2 = Ek ./(pn + 0.001); % Normalized error

pn是當前幀遠端信號功率譜，Ek是誤差信號頻譜。Ek2是歸一化誤差頻譜。NLMS公式要求。

    absEf = max(abs(Ek2), threshold);
    absEf = ones(N+1,1)*threshold./absEf;
    Ek2 = Ek2.*absEf;

max的作用是爲了防止歸一化後誤差頻譜過小，最終得到的Ek2是一個限幅矩陣，如果該點的值比門限大，則取門限，如果該點的值比門限小，則保持不變。

 mEk = mufb.*Ek2;

mufb是步長，對於16000情況，步長取了0.8.NLMS公式。

 PP = conj(XFm).*(ones(M,1) * mEk')';
    tmp = [PP ; flipud(conj(PP(2:N,:)))];
    IFPP = real(ifft(tmp));
    PH = IFPP(1:N,:);
    tmp = fft([PH;zeros(N,M)]);
    FPH = tmp(1:N+1,:);
    WFb = WFb + FPH;

PP是將遠端信號的共軛乘以誤差信號頻譜，這一項用於調節步長，NLMS（步長=參考信號×步長×誤差）的可變步長就提現在這裏。PH是頻域到時域的變換值。這和前面頻域到時域的變換原理一樣。WFb是權中係數的更新。

    if mod(kk, 10*mult) == 0
        WFbEn = sum(real(WFb.*conj(WFb)));
        %WFbEn = sum(abs(WFb));
        [tmp, dIdx] = max(WFbEn);

        WFbD = sum(abs(WFb(:, dIdx)),2);
        %WFbD = WFbD / (mean(WFbD) + 1e-10);
        WFbD = min(max(WFbD, 0.5), 4);
    end
     dIdxV(kk) = dIdx;

上述的作用是更新dIdx和dIdxV。這裏的更新並不是每一次都更新，一來是爲了穩定，而來也是變相的減少計算量，提高實時性。就算是每一次都更新dIdx，WebRTC計算速度評估的結果也是很滿意的。WFb是權重向量的頻譜表示，WFbEn是權重向量按列求和，得到的是16*1的矩陣。這樣得到的是16個塊對權重的累加和。這樣的區分度比直接累加和要大。
[tmp, dIdx] = max(WFbEn);作用就是找到16個塊中權重累加和最大值及其對應的索引。
WFbD首先計算了權重最大那個塊dIdx的列，然後將其按行求和，得到的就是65*1矩陣。WFbD不能低於0.5也不能高於4，算法中並未使用到，plot性能分析時用到。
最後把索引值dIdx存放到dIdxV(kk)中，這樣每來一幀，就會有一個最大索引值放到dIdxV向量中。

功率譜密度和相關性計算

NLP

這裏的NLP不是native language processing，而是Non-linear processing的意思。

        ee = [eo;ekfb];
        eo = ekfb;
        window = wins;

上述作用是將上次的誤差和ekfb組合，其中eo可以理解爲error old。eo也確實保存了上一次的誤差。window是簡單將窗函數賦值。

        tmp = fft(xx.*window);
        xf = tmp(1:N+1);
        tmp = fft(dd.*window);
        df = tmp(1:N+1);
        tmp = fft(ee.*window);
        ef = tmp(1:N+1);

上述代碼是把xx，dd，ee加窗後做FFT變換，並且取了fs/2fs/2的頻譜部分存放到xf，df以及ef中。加窗的目的是爲了減小頻譜泄露，提高譜分辨率。

        xfwm(:,1) = xf;
        xf = xfwm(:,dIdx);
        %fprintf(1,'%d: %f\n', kk, xf(4));
        dfm(:,1) = df;

將xf存放到xfwm的第一列，xfwm是65*16的矩陣，df做類似處理。
然後把dIdx指向的那一列傳給xf，dIdx是之前計算的權重矩陣能量最大的那塊的索引，這樣從xfwm矩陣裏把真正要處理近端信號對應的遠端信號（speaker，參考信號）獲取到。

        Se = gamma*Se + (1-gamma)*real(ef.*conj(ef));
        Sd = gamma*Sd + (1-gamma)*real(df.*conj(df));
        Sx = gamma*Sx + (1 - gamma)*real(xf.*conj(xf));

計算ef，df和xf的平滑功率譜，gamma這裏取值是0.92.相對於8k信號取值略大。它們都是65*1的矩陣，包括了直流分量的能力，剩下的64點是fs/2fs/2及以下的頻點能量。

        Sxd = gamma*Sxd + (1 - gamma)*xf.*conj(df);
        Sed = gamma*Sed + (1-gamma)*ef.*conj(df);

計算互功率譜，這裏計算了遠端信號和近端信號功率譜，如果該值較大，則說明它們的相關性較強，說明近端信號採集到了遠端信號的概率很大，這時需要進行噪聲抑制，同樣的如果誤差信號和近端信號功率譜較大，則說明估計的y^y^是比較準的，誤差信號裏剩餘的遠端信號較少，需要進行噪聲抑制的概率就小。它們也都是65*1矩陣，對應頻點的bin。但是上述獲得的是絕對值而非相對值，門限設置不容易，需要一個歸一化的過程。歸一化的過程可以通過求互相關得到。

        cohed = real(Sed.*conj(Sed))./(Se.*Sd + 1e-10);
        cohedAvg(kk) = mean(cohed(echoBandRange));
        cohxd = real(Sxd.*conj(Sxd))./(Sx.*Sd + 1e-10);

如上，計算誤差信號和近端信號的互相關，1e-10是爲了防止除0報錯。cohed越大，表示回聲越小，cohxd越大，表示回聲越大，這裏就可以設置一個統一的門限評判上下限了。

cohedMean = mean(cohed(echoBandRange));

計算設置的echoBandRange裏頻點的均值，如果echoBandRange設置的過大，則低音部分如鼓點聲則不易消掉。

        hnled = min(1 - cohxd, cohed);

這裏的作用就是把最小值賦值給hnled，該值越大，則說明越不需要消回聲。之所以取二者判斷，是爲了最大可能性的消除掉回聲。

        [hnlSort,   hnlSortIdx] = sort(1-cohxd(echoBandRange));
        [xSort, xSortIdx] = sort(Sx);

對1-cohxd（echoBandRange）進行升序排序，同樣對Sx也進行升序排序。

hnlSortQ = mean(1 - cohxd(echoBandRange));

對遠端和近端信號的帶內互相關求均值。hnlSortQ表示遠端和近端不相關性的平均值，其值越大約沒有回聲，與cohed含義一致。

 [hnlSort2, hnlSortIdx2] = sort(hnled(echoBandRange));

對hnled進行升序排序。

        hnlQuant = 0.75;
        hnlQuantLow = 0.5;
        qIdx = floor(hnlQuant*length(hnlSort2));
        qIdxLow = floor(hnlQuantLow*length(hnlSort2));
        hnlPrefAvg = hnlSort2(qIdx);
        hnlPrefAvgLow = hnlSort2(qIdxLow);

這裏主要取了兩個值，一個值取在了排序後的3/4處，一個值取在了排序後的1/2處。

            if cohedMean > 0.98 & hnlSortQ > 0.9
                suppState = 0;
            elseif cohedMean < 0.95 | hnlSortQ < 0.8
                suppState = 1;
            end

如果誤差和近端信號的互相關均值大於0.98，且遠端和近端頻帶內的互不相關大於0.9，則說明不需要進行回聲抑制，將suppState標誌設置成0，如果誤差和近端信號小於0.95或者遠端和近端頻帶內信號不相關性小於0.8則需要進行印製，在這個範圍之外的，回聲抑制標誌保持前一幀的狀態。

            if hnlSortQ < cohxdLocalMin & hnlSortQ < 0.75
                cohxdLocalMin = hnlSortQ;
            end

cohxdLocalMin的初始值是1，表示遠端和近端完全不相關，這裏判斷計算得到的遠端近端不相關性是否小於前一次的不相關性，如果小於且不相關性小於0.75，則更新cohxdLocalMin。

            if cohxdLocalMin == 1
                ovrd = 3;
                hnled = 1-cohxd;
                hnlPrefAvg = hnlSortQ;
                hnlPrefAvgLow = hnlSortQ;
            end

如果cohxdLocalMin=1，則說明要麼是發現遠端和近端完全不相關，要麼就是cohxdLocalMin一直沒有更新，既然不相關性很大，那麼也說明有回聲的可能性小，那麼使用較小的ovrd（over-driven）值，和較大的hnled（65*1）值。

            if suppState == 0
                hnled = cohed;
                hnlPrefAvg = cohedMean;
                hnlPrefAvgLow = cohedMean;
            end

如果suppState==0，則說明不需要進行回聲消除，直接用誤差近端相關性修正誤差信號，hnl的兩個均值讀取cohed的均值，在這種情況下hnled的值接近於1.

            if hnlPrefAvgLow < hnlLocalMin & hnlPrefAvgLow < 0.6
                hnlLocalMin = hnlPrefAvgLow;
                hnlMin = hnlPrefAvgLow;
                hnlNewMin = 1;
                hnlMinCtr = 0;
                if hnlMinCtr == 0
                    hnlMinCtr = hnlMinCtr + 1;
                else
                    hnlMinCtr = 0;
                    hnlMin = hnlLocalMin;
                    SeLocalMin = SeQ;
                    SdLocalMin = SdQ;
                    SeLocalAvg = 0;
                    minCtr = 0;
                    ovrd = max(log(0.0001)/log(hnlMin), 2);
                    divergeFact = hnlLocalMin;
                end
            end

hnlLocalMin是hnl係數的最小值，在滿足這條判斷的情況下發現了更小的值，需要對其進行更新，同時表標誌設置成1，計數清0，這種情況下回聲的概率較大，所以把ovrd設大以增強抑制能力。

            if hnlMinCtr == 2
                hnlNewMin = 0;
                hnlMinCtr = 0;
                ovrd = max(log(0.00000001)/(log(hnlMin + 1e-10) + 1e-10), 5);

            end

hnlMinCtr==2，說明之前有滿足<0.6的塊使得hnlMinCtr=2，然後其下一塊又沒有滿足<0.6的條件，進而更新了ovrd值。默認是和3比較取最大值，這裏調節成5是爲了增加抑制效果，實際情況還需要針對系統調試。

            hnlLocalMin = min(hnlLocalMin + 0.0008/mult, 1);
            cohxdLocalMin = min(cohxdLocalMin + 0.0004/mult, 1);

跟新上述兩個值，mult是fs/8000fs/8000,對於16kHz，就是2.就是0.0004和0.0002的差異。

            if ovrd < ovrdSm
                ovrdSm = 0.99*ovrdSm + 0.01*ovrd;
            else
                ovrdSm = 0.9*ovrdSm + 0.1*ovrd;
            end

平滑更新ovrdSm，上述結果傾向於保持ovrdSm是一個較大的值，這個較大的值是爲了儘量抑制回聲。

        ekEn = sum(Se);
        dkEn = sum(Sd);

按行求和，物理意義分別是誤差能量和近端信號能量。

發散處理

 if divergeState == 0
            if ekEn > dkEn
                ef = df;
                divergeState = 1;
            end
        else
            if ekEn*1.05 < dkEn
                divergeState = 0;
            else
                ef = df;
            end
        end

如果不進行發散處理，誤差能量大於近端能力，則用近端頻譜更新誤差頻譜並把發散狀態設置成1，如果誤差能量的1.05倍小於近端能量，則發散處理標誌設置成0.

        if ekEn > dkEn*19.95
            WFb=zeros(N+1,M); % Block-based FD NLMS
        end

如果誤差能量大於近端能量×19.95則，將權重係數矩陣設置成0.

        ekEnV(kk) = ekEn;
        dkEnV(kk) = dkEn;

相應能量存放在相應的向量中。

        hnlLocalMinV(kk) = hnlLocalMin;
        cohxdLocalMinV(kk) = cohxdLocalMin;
        hnlMinV(kk) = hnlMin;

上述三個向量保存對應值。

平滑濾波器係數和抑制指數

        wCurve = [0; aggrFact*sqrt(linspace(0,1,N))' + 0.1];

權重係數曲線生成，線性均勻分佈。

    hnled = weight.*min(hnlPrefAvg, hnled) + (1 - weight).*hnled;

使用權重平滑hnled，wCurve分佈是讓65點中頻率越高的點本次hnled佔的比重越大，反之則以前的平滑結果所佔比重大。

od = ovrdSm*(sqrt(linspace(0,1,N+1))' + 1);

產生65*1的曲線，用作更新hnled的冪指數。

      hnled = hnled.^(od.*sshift);

od作爲冪指數跟新hnled。

輸出回聲消除後的信號

 hnl = hnled;
 ef = ef.*(hnl);

用hnl係數與誤差頻譜相乘，即頻域卷積，就是將誤差信號通過了傳遞函數爲hnnl的濾波器。

        ovrdV(kk) = ovrdSm;
        hnledAvg(kk) = 1-mean(1-cohed(echoBandRange));
        hnlxdAvg(kk) = 1-mean(cohxd(echoBandRange));
        hnlSortQV(kk) = hnlPrefAvgLow;
        hnlPrefAvgV(kk) = hnlPrefAvg;

相關值的暫存
沒有開啓舒適噪聲產生，則Fmix=ef。

    % Overlap and add in time domain for smoothness
    tmp = [Fmix ; flipud(conj(Fmix(2:N)))];
    mixw = wins.*real(ifft(tmp));
    mola = mbuf(end-N+1:end) + mixw(1:N);
    mbuf = mixw;
    ercn(pos:pos+N-1) = mola;

則使用重疊想加法獲得時域平滑信號。

    XFm(:,2:end) = XFm(:,1:end-1);
    YFm(:,2:end) = YFm(:,1:end-1);
    xfwm(:,2:end) = xfwm(:,1:end-1);
    dfm(:,2:end) = dfm(:,1:end-1);

全體後移一個塊，進入下一塊迭代。
執行完了之後，如果想聽回聲消除後信號的聲音，使用如下命令：
sound(10*ercn,16000)
其中16000是輸入信號的頻率。

整體的Matlab代碼如下：

% Partitioned block frequency domain adaptive filtering NLMS and
% standard time-domain sample-based NLMS
%near is micphone captured signal
fid=fopen('near.pcm', 'rb'); % Load far end
ssin=fread(fid,inf,'float32');
fclose(fid);
%far is speaker played music
fid=fopen('far.pcm', 'rb'); % Load fnear end
rrin=fread(fid,inf,'float32');
fclose(fid);

rand('state',13);
fs=16000;
mult=fs/8000;
if fs == 8000
cohRange = 2:3;
elseif fs==16000
cohRange = 2;
end

% Flags
NLPon=1; % NLP on
CNon=0; % Comfort noise on
PLTon=0; % Plotting on

M = 16; % Number of partitions
N = 64; % Partition length
L = M*N; % Filter length
if fs == 8000
    mufb = 0.6;
else
    mufb = 0.8;
end
VADtd=48;
alp = 0.15; % Power estimation factor 
alc = 0.1; % Coherence estimation factor
beta = 0.9; % Plotting factor
%% Changed a little %%
step = 0.1875;%0.1875; % Downward step size
%%
if fs == 8000
    threshold=2e-6; % DTrob threshold
else
    %threshold=0.7e-6;
    threshold=1.5e-6; 
end

if fs == 8000
    echoBandRange = ceil(300*2/fs*N):floor(1800*2/fs*N);
else
    echoBandRange = ceil(60*2/fs*N):floor(1500*2/fs*N);
end
suppState = 1;
transCtr = 0;

Nt=1;
vt=1;

ramp = 1.0003; % Upward ramp
rampd = 0.999; % Downward ramp
cvt = 20; % Subband VAD threshold;
nnthres = 20; % Noise threshold

shh=logspace(-1.3,-2.2,N+1)';
sh=[shh;flipud(shh(2:end-1))]; % Suppression profile

len=length(ssin);
w=zeros(L,1); % Sample-based TD(time domain) NLMS
WFb=zeros(N+1,M); % Block-based FD(frequency domain) NLMS
WFbOld=zeros(N+1,M); % Block-based FD NLMS
YFb=zeros(N+1,M);
erfb=zeros(len,1);
erfb3=zeros(len,1);

ercn=zeros(len,1);
zm=zeros(N,1);
XFm=zeros(N+1,M);
YFm=zeros(N+1,M);
pn0=10*ones(N+1,1);
pn=zeros(N+1,1);
NN=len;
Nb=floor(NN/N)-M;
erifb=zeros(Nb+1,1)+0.1;
erifb3=zeros(Nb+1,1)+0.1;
ericn=zeros(Nb+1,1)+0.1;
dri=zeros(Nb+1,1)+0.1;
start=1;
xo=zeros(N,1);
do=xo;
eo=xo;

echoBands=zeros(Nb+1,1);
cohxdAvg=zeros(Nb+1,1);
cohxdSlow=zeros(Nb+1,N+1);
cohedSlow=zeros(Nb+1,N+1);
%overdriveM=zeros(Nb+1,N+1);
cohxdFastAvg=zeros(Nb+1,1);
cohxdAvgBad=zeros(Nb+1,1);
cohedAvg=zeros(Nb+1,1);
cohedFastAvg=zeros(Nb+1,1);
hnledAvg=zeros(Nb+1,1);
hnlxdAvg=zeros(Nb+1,1);
ovrdV=zeros(Nb+1,1);
dIdxV=zeros(Nb+1,1);
SLxV=zeros(Nb+1,1);
hnlSortQV=zeros(Nb+1,1);
hnlPrefAvgV=zeros(Nb+1,1);
mutInfAvg=zeros(Nb+1,1);
%overdrive=zeros(Nb+1,1);
hnled = zeros(N+1, 1);
weight=zeros(N+1,1);
hnlMax = zeros(N+1, 1);
hnl = zeros(N+1, 1);
overdrive = ones(1, N+1);
xfwm=zeros(N+1,M);
dfm=zeros(N+1,M);
WFbD=ones(N+1,1);

fbSupp = 0;
hnlLocalMin = 1;
cohxdLocalMin = 1;
hnlLocalMinV=zeros(Nb+1,1);
cohxdLocalMinV=zeros(Nb+1,1);
hnlMinV=zeros(Nb+1,1);
dkEnV=zeros(Nb+1,1);
ekEnV=zeros(Nb+1,1);
ovrd = 2;
ovrdPos = floor((N+1)/4);
ovrdSm = 2;
hnlMin = 1;
minCtr = 0;
SeMin = 0;
SdMin = 0;
SeLocalAvg = 0;
SeMinSm = 0;
divergeFact = 1;
dIdx = 1;
hnlMinCtr = 0;
hnlNewMin = 0;
divergeState = 0;

Sy=ones(N+1,1);
Sym=1e7*ones(N+1,1);

wins=[0;sqrt(hanning(2*N-1))];
ubufn=zeros(2*N,1);
ebuf=zeros(2*N,1);
ebuf2=zeros(2*N,1);
ebuf4=zeros(2*N,1);
mbuf=zeros(2*N,1);

cohedFast = zeros(N+1,1);
cohxdFast = zeros(N+1,1);
cohxd = zeros(N+1,1);
Se = zeros(N+1,1);
Sd = zeros(N+1,1);
Sx = zeros(N+1,1);
SxBad = zeros(N+1,1);
Sed = zeros(N+1,1);
Sxd = zeros(N+1,1);
SxdBad = zeros(N+1,1);
hnledp=[];

cohxdMax = 0;

hh=waitbar(0,'Please wait...');
%progressbar(0);

%spaces = ' ';
%spaces = repmat(spaces, 50, 1);
%spaces = ['[' ; spaces ; ']'];
%fprintf(1, spaces);
%fprintf(1, '\n');

for kk=1:Nb
    pos = N * (kk-1) + start;

    % FD block method
    % ---------------------- Organize data

    %far is speaker played music
    xk = rrin(pos:pos+N-1);
    %near is micphone captured signal
    dk = ssin(pos:pos+N-1);

    %----------------------- far end signal process
    xx = [xo;xk];
    xo = xk;
    tmp = fft(xx);
    XX = tmp(1:N+1);

    dd = [do;dk]; % Overlap
    do = dk;
    tmp = fft(dd); % Frequency domain
    DD = tmp(1:N+1);

    % ------------------------far end Power estimation
    pn0 = (1 - alp) * pn0 + alp * real(XX.* conj(XX));
    pn = pn0;
%   pn = (1 - alp) * pn + alp * M * pn0;

    % ---------------------- Filtering
    XFm(:,1) = XX;
    for mm=0:(M-1)
        m=mm+1;
        YFb(:,m) = XFm(:,m) .* WFb(:,m);
    end
    yfk = sum(YFb,2);
    tmp = [yfk ; flipud(conj(yfk(2:N)))];
    ykt = real(ifft(tmp));
    ykfb = ykt(end-N+1:end);

    % ---------------------- Error estimation
    ekfb = dk - ykfb;
    %if sum(abs(ekfb)) < sum(abs(dk))
        %ekfb = dk - ykfb;
    % erfb(pos:pos+N-1) = ekfb;
    %else
        %ekfb = dk;
    % erfb(pos:pos+N-1) = dk;
    %end
%(kk-1)*(N*2)+1
    erfb(pos:pos+N-1) = ekfb;
    tmp = fft([zm;ekfb]); % FD version for cancelling part (overlap-save)
    Ek = tmp(1:N+1);

    % ------------------------ Adaptation
    %Ek2 = Ek ./(M*pn + 0.001); % Normalized error
    Ek2 = Ek ./(pn + 0.001); % Normalized error

    absEf = max(abs(Ek2), threshold);
    absEf = ones(N+1,1)*threshold./absEf;
    Ek2 = Ek2.*absEf;

    mEk = mufb.*Ek2;
    PP = conj(XFm).*(ones(M,1) * mEk')';
    tmp = [PP ; flipud(conj(PP(2:N,:)))];
    IFPP = real(ifft(tmp));
    PH = IFPP(1:N,:);
    tmp = fft([PH;zeros(N,M)]);
    FPH = tmp(1:N+1,:);
    WFb = WFb + FPH;

%     if mod(kk, 10*mult) == 0
        WFbEn = sum(real(WFb.*conj(WFb)));
        %WFbEn = sum(abs(WFb));
        [tmp, dIdx] = max(WFbEn);

        WFbD = sum(abs(WFb(:, dIdx)),2);
        %WFbD = WFbD / (mean(WFbD) + 1e-10);
        WFbD = min(max(WFbD, 0.5), 4);
%     end
    dIdxV(kk) = dIdx;

    % NLP
    if (NLPon)  
        ee = [eo;ekfb];
        eo = ekfb;
        window = wins;
        if fs == 8000
            gamma = 0.9;
        else
        gamma = 0.93;
        end

        tmp = fft(xx.*window);
        xf = tmp(1:N+1);
        tmp = fft(dd.*window);
        df = tmp(1:N+1);
        tmp = fft(ee.*window);
        ef = tmp(1:N+1);

        xfwm(:,1) = xf;
        xf = xfwm(:,dIdx);
        %fprintf(1,'%d: %f\n', kk, xf(4));
        dfm(:,1) = df;

        SxOld = Sx;

        Se = gamma*Se + (1-gamma)*real(ef.*conj(ef));
        Sd = gamma*Sd + (1-gamma)*real(df.*conj(df));
        Sx = gamma*Sx + (1 - gamma)*real(xf.*conj(xf));

        %xRatio = real(xfwm(:,1).*conj(xfwm(:,1))) ./ ...
        % (real(xfwm(:,2).*conj(xfwm(:,2))) + 1e-10);
        %xRatio = Sx ./ (SxOld + 1e-10);
        %SLx = log(1/(N+1)*sum(xRatio)) - 1/(N+1)*sum(log(xRatio));
        %SLxV(kk) = SLx;

%         freqSm = 0.9;
%         Sx = filter(freqSm, [1 -(1-freqSm)], Sx);
%         Sx(end:1) = filter(freqSm, [1 -(1-freqSm)], Sx(end:1));
%         Se = filter(freqSm, [1 -(1-freqSm)], Se);
%         Se(end:1) = filter(freqSm, [1 -(1-freqSm)], Se(end:1));
%         Sd = filter(freqSm, [1 -(1-freqSm)], Sd);
%         Sd(end:1) = filter(freqSm, [1 -(1-freqSm)], Sd(end:1));

        %SeFast = ef.*conj(ef);
        %SdFast = df.*conj(df);
        %SxFast = xf.*conj(xf);
        %cohedFast = 0.9*cohedFast + 0.1*SeFast ./ (SdFast + 1e-10);
        %cohedFast(find(cohedFast > 1)) = 1;
        %cohedFast(find(cohedFast > 1)) = 1 ./ cohedFast(find(cohedFast>1));
        %cohedFastAvg(kk) = mean(cohedFast(echoBandRange));
        %cohedFastAvg(kk) = min(cohedFast);

        %cohxdFast = 0.8*cohxdFast + 0.2*log(SdFast ./ (SxFast + 1e-10));
        %cohxdFastAvg(kk) = mean(cohxdFast(echoBandRange));

        % coherence
        Sxd = gamma*Sxd + (1 - gamma)*xf.*conj(df);
        Sed = gamma*Sed + (1-gamma)*ef.*conj(df);

%         Sxd = filter(freqSm, [1 -(1-freqSm)], Sxd);
%         Sxd(end:1) = filter(freqSm, [1 -(1-freqSm)], Sxd(end:1));
%         Sed = filter(freqSm, [1 -(1-freqSm)], Sed);
%         Sed(end:1) = filter(freqSm, [1 -(1-freqSm)], Sed(end:1));

        cohed = real(Sed.*conj(Sed))./(Se.*Sd + 1e-10);
        cohedAvg(kk) = mean(cohed(echoBandRange));
        %cohedAvg(kk) = cohed(6);
        %cohedAvg(kk) = min(cohed);

        cohxd = real(Sxd.*conj(Sxd))./(Sx.*Sd + 1e-10);
        freqSm = 0.6;
        cohxd(2:end) = filter(freqSm, [1 -(1-freqSm)], cohxd(2:end));
        cohxd(end:2) = filter(freqSm, [1 -(1-freqSm)], cohxd(end:2));
        cohxdAvg(kk) = mean(cohxd(echoBandRange));
        %cohxdAvg(kk) = (cohxd(32));
        %cohxdAvg(kk) = max(cohxd);

        %xf = xfm(:,dIdx);
        %SxBad = gamma*SxBad + (1 - gamma)*real(xf.*conj(xf));
        %SxdBad = gamma*SxdBad + (1 - gamma)*xf.*conj(df);
        %cohxdBad = real(SxdBad.*conj(SxdBad))./(SxBad.*Sd + 0.01);
        %cohxdAvgBad(kk) = mean(cohxdBad);

        %for j=1:N+1
        % mutInf(j) = 0.9*mutInf(j) + 0.1*information(abs(xfm(j,:)), abs(dfm(j,:)));
        %end
        %mutInfAvg(kk) = mean(mutInf);

        %hnled = cohedFast;
        %xIdx = find(cohxd > 1 - cohed);
        %hnled(xIdx) = 1 - cohxd(xIdx);
        %hnled = 1 - max(cohxd, 1-cohedFast);
        hnled = min(1 - cohxd, cohed);
        %hnled = 1 - cohxd;
        %hnled = max(1 - (cohxd + (1-cohedFast)), 0);
        %hnled = 1 - max(cohxd, 1-cohed);

        if kk > 1
            cohxdSlow(kk,:) = 0.99*cohxdSlow(kk-1,:) + 0.01*cohxd';
            cohedSlow(kk,:) = 0.99*cohedSlow(kk-1,:) + 0.01*(1-cohed)';
        end


        if 0
        %if kk > 50
            %idx = find(hnled > 0.3);
            hnlMax = hnlMax*0.9999;
            %hnlMax(idx) = max(hnlMax(idx), hnled(idx));
            hnlMax = max(hnlMax, hnled);
            %overdrive(idx) = max(log(hnlMax(idx))/log(0.99), 1);
            avgHnl = mean(hnlMax(echoBandRange));
            if avgHnl > 0.3
                overdrive = max(log(avgHnl)/log(0.99), 1);
            end
            weight(4:end) = max(hnlMax) - hnlMax(4:end);
        end



        %[hg, gidx] = max(hnled);
        %fnrg = Sx(gidx) / (Sd(gidx) + 1e-10);

        %[tmp, bidx] = find((Sx / Sd + 1e-10) > fnrg);
        %hnled(bidx) = hg;


        %cohed1 = mean(cohed(cohRange)); % range depends on bandwidth
        %cohed1 = cohed1^2;
        %echoBands(kk) = length(find(cohed(echoBandRange) < 0.25))/length(echoBandRange);

        %if (fbSupp == 0)
        % if (echoBands(kk) > 0.8)
        % fbSupp = 1;
        % end
        %else
        % if (echoBands(kk) < 0.6)
        % fbSupp = 0;
        % end
        %end
        %overdrive(kk) = 7.5*echoBands(kk) + 0.5;

% Factor by which to weight other bands
%if (cohed1 < 0.1)
% w = 0.8 - cohed1*10*0.4;
%else
% w = 0.4;
%end

% Weight coherence subbands
%hnled = w*cohed1 + (1 - w)*cohed;
%hnled = (hnled).^2;
%cohed(floor(N/2):end) = cohed(floor(N/2):end).^2;
        %if fbSupp == 1
        % cohed = zeros(size(cohed));
        %end
        %cohed = cohed.^overdrive(kk);

        %hnled = gamma*hnled + (1 - gamma)*cohed;
% Additional hf suppression
%hnledp = [hnledp ; mean(hnled)];
%hnled(floor(N/2):end) = hnled(floor(N/2):end).^2;
%ef = ef.*((weight*(min(1 - hnled)).^2 + (1 - weight).*(1 - hnled)).^2);

        cohedMean = mean(cohed(echoBandRange));
        %aggrFact = 4*(1-mean(hnled(echoBandRange))) + 1;
        %[hnlSort, hnlSortIdx] = sort(hnled(echoBandRange));
        [hnlSort,   hnlSortIdx] = sort(1-cohxd(echoBandRange));
        [xSort, xSortIdx] = sort(Sx);
        %aggrFact = (1-mean(hnled(echoBandRange)));
        %hnlSortQ = hnlSort(qIdx);
        hnlSortQ = mean(1 - cohxd(echoBandRange));
        %hnlSortQ = mean(1 - cohxd);

        [hnlSort2, hnlSortIdx2] = sort(hnled(echoBandRange));
        %[hnlSort2, hnlSortIdx2] = sort(hnled);
        hnlQuant = 0.75;
        hnlQuantLow = 0.5;
        qIdx = floor(hnlQuant*length(hnlSort2));
        qIdxLow = floor(hnlQuantLow*length(hnlSort2));
        hnlPrefAvg = hnlSort2(qIdx);
        hnlPrefAvgLow = hnlSort2(qIdxLow);
        %hnlPrefAvgLow = mean(hnled);
        %hnlPrefAvg = max(hnlSort2);
        %hnlPrefAvgLow = min(hnlSort2);

        %hnlPref = hnled(echoBandRange);
        %hnlPrefAvg = mean(hnlPref(xSortIdx((0.5*length(xSortIdx)):end)));

        %hnlPrefAvg = min(hnlPrefAvg, hnlSortQ);

        %hnlSortQIdx = hnlSortIdx(qIdx);
        %SeQ = Se(qIdx + echoBandRange(1) - 1);
        %SdQ = Sd(qIdx + echoBandRange(1) - 1);
        %SeQ = Se(qIdxLow + echoBandRange(1) - 1);
        %SdQ = Sd(qIdxLow + echoBandRange(1) - 1);
        %propLow = length(find(hnlSort < 0.1))/length(hnlSort);
        %aggrFact = min((1 - hnlSortQ)/2, 0.5);
        %aggrTerm = 1/aggrFact;

        %hnlg = mean(hnled(echoBandRange));
        %hnlg = hnlSortQ;
        %if suppState == 0
        % if hnlg < 0.05
        % suppState = 2;
        % transCtr = 0;
        % elseif hnlg < 0.75
        % suppState = 1;
        % transCtr = 0;
        % end
        %elseif suppState == 1
        % if hnlg > 0.8
        % suppState = 0;
        % transCtr = 0;
        % elseif hnlg < 0.05
        % suppState = 2;
        % transCtr = 0;
        % end
        %else
        % if hnlg > 0.8
        % suppState = 0;
        % transCtr = 0;
        % elseif hnlg > 0.25
        % suppState = 1;
        % transCtr = 0;
        % end
        %end
        %if kk > 50

            if cohedMean > 0.98 & hnlSortQ > 0.9
                %if suppState == 1
                % hnled = 0.5*hnled + 0.5*cohed;
                % %hnlSortQ = 0.5*hnlSortQ + 0.5*cohedMean;
                % hnlPrefAvg = 0.5*hnlPrefAvg + 0.5*cohedMean;
                %else
                % hnled = cohed;
                % %hnlSortQ = cohedMean;
                % hnlPrefAvg = cohedMean;
                %end
                suppState = 0;
            elseif cohedMean < 0.95 | hnlSortQ < 0.8
                %if suppState == 0
                % hnled = 0.5*hnled + 0.5*cohed;
                % %hnlSortQ = 0.5*hnlSortQ + 0.5*cohedMean;
                % hnlPrefAvg = 0.5*hnlPrefAvg + 0.5*cohedMean;
                %end
                suppState = 1;
            end

            if hnlSortQ < cohxdLocalMin & hnlSortQ < 0.75
                cohxdLocalMin = hnlSortQ;
            end

            if cohxdLocalMin == 1
                ovrd = 3;
                hnled = 1-cohxd;
                hnlPrefAvg = hnlSortQ;
                hnlPrefAvgLow = hnlSortQ;
            end

            if suppState == 0
                hnled = cohed;
                hnlPrefAvg = cohedMean;
                hnlPrefAvgLow = cohedMean;
            end

            %if hnlPrefAvg < hnlLocalMin & hnlPrefAvg < 0.6
            if hnlPrefAvgLow < hnlLocalMin & hnlPrefAvgLow < 0.6
                %hnlLocalMin = hnlPrefAvg;
                %hnlMin = hnlPrefAvg;
                hnlLocalMin = hnlPrefAvgLow;
                hnlMin = hnlPrefAvgLow;
                hnlNewMin = 1;
                hnlMinCtr = 0;
                if hnlMinCtr == 0
                    hnlMinCtr = hnlMinCtr + 1;
                else
                    hnlMinCtr = 0;
                    hnlMin = hnlLocalMin;
                    SeLocalMin = SeQ;
                    SdLocalMin = SdQ;
                    SeLocalAvg = 0;
                    minCtr = 0;
                    ovrd = max(log(0.0001)/log(hnlMin), 2);
                    divergeFact = hnlLocalMin;
                end
            end

            if hnlNewMin == 1
                hnlMinCtr = hnlMinCtr + 1;
            end
            if hnlMinCtr == 2
                hnlNewMin = 0;
                hnlMinCtr = 0;
                %ovrd = max(log(0.0001)/log(hnlMin), 2);
%                 ovrd = max(log(0.00001)/(log(hnlMin + 1e-10) + 1e-10), 3);
                ovrd = max(log(0.00000001)/(log(hnlMin + 1e-10) + 1e-10), 5);
                %ovrd = max(log(0.0001)/log(hnlPrefAvg), 2);
                %ovrd = max(log(0.001)/log(hnlMin), 2);
            end
            hnlLocalMin = min(hnlLocalMin + 0.0008/mult, 1);
            cohxdLocalMin = min(cohxdLocalMin + 0.0004/mult, 1);
            %divergeFact = hnlSortQ;


            %if minCtr > 0 & hnlLocalMin < 1
            % hnlMin = hnlLocalMin;
            % %SeMin = 0.9*SeMin + 0.1*sqrt(SeLocalMin);
            % SdMin = sqrt(SdLocalMin);
            % %SeMin = sqrt(SeLocalMin)*hnlSortQ;
            % SeMin = sqrt(SeLocalMin);
            % %ovrd = log(100/SeMin)/log(hnlSortQ);
            % %ovrd = log(100/SeMin)/log(hnlSortQ);
            % ovrd = log(0.01)/log(hnlMin);
            % ovrd = max(ovrd, 2);
            % ovrdPos = hnlSortQIdx;
            % %ovrd = max(ovrd, 1);
            % %SeMin = sqrt(SeLocalAvg/5);
            % minCtr = 0;
            %else
            % %SeLocalMin = 0.9*SeLocalMin +0.1*SeQ;
            % SeLocalAvg = SeLocalAvg + SeQ;
            % minCtr = minCtr + 1;
            %end

            if ovrd < ovrdSm
                ovrdSm = 0.99*ovrdSm + 0.01*ovrd;
            else
                ovrdSm = 0.9*ovrdSm + 0.1*ovrd;
            end
        %end

%         ekEn = sum(real(ekfb.^2));
%         dkEn = sum(real(dk.^2));
        ekEn = sum(Se);
        dkEn = sum(Sd);

        if divergeState == 0
            if ekEn > dkEn
                ef = df;
                divergeState = 1;
                %hnlPrefAvg = hnlSortQ;
                %hnled = (1 - cohxd);
            end
        else
            %if ekEn*1.1 < dkEn
            %if ekEn*1.26 < dkEn
            if ekEn*1.05 < dkEn
                divergeState = 0;
            else
                ef = df;
            end
        end

        if ekEn > dkEn*19.95
            WFb=zeros(N+1,M); % Block-based FD NLMS
        end

        ekEnV(kk) = ekEn;
        dkEnV(kk) = dkEn;

        hnlLocalMinV(kk) = hnlLocalMin;
        cohxdLocalMinV(kk) = cohxdLocalMin;
        hnlMinV(kk) = hnlMin;
        %cohxdMaxLocal = max(cohxdSlow(kk,:));
        %if kk > 50
        %cohxdMaxLocal = 1-hnlSortQ;
        %if cohxdMaxLocal > 0.5
        % %if cohxdMaxLocal > cohxdMax
        % odScale = max(log(cohxdMaxLocal)/log(0.95), 1);
        % %overdrive(7:end) = max(log(cohxdSlow(kk,7:end))/log(0.9), 1);
        % cohxdMax = cohxdMaxLocal;
        % end
        %end
        %end
        %cohxdMax = cohxdMax*0.999;

        %overdriveM(kk,:) = max(overdrive, 1);
        %aggrFact = 0.25;
        aggrFact = 0.3;
        %aggrFact = 0.5*propLow;
        %if fs == 8000
        % wCurve = [0 ; 0 ; aggrFact*sqrt(linspace(0,1,N-1))' + 0.1];
        %else
        % wCurve = [0; 0; 0; aggrFact*sqrt(linspace(0,1,N-2))' + 0.1];
        %end
        wCurve = [0; aggrFact*sqrt(linspace(0,1,N))' + 0.1];
        % For sync with C
        %if fs == 8000
        % wCurve = wCurve(2:end);
        %else
        % wCurve = wCurve(1:end-1);
        %end
        %weight = aggrFact*(sqrt(linspace(0,1,N+1)'));
        %weight = aggrFact*wCurve;
        weight = wCurve;
        %weight = aggrFact*ones(N+1,1);
        %weight = zeros(N+1,1);
        %hnled = weight.*min(hnled) + (1 - weight).*hnled;
        %hnled = weight.*min(mean(hnled(echoBandRange)), hnled) + (1 - weight).*hnled;
        %hnled = weight.*min(hnlSortQ, hnled) + (1 - weight).*hnled;

        %hnlSortQV(kk) = mean(hnled);
        %hnlPrefAvgV(kk) = mean(hnled(echoBandRange));

        hnled = weight.*min(hnlPrefAvg, hnled) + (1 - weight).*hnled;

        %od = aggrFact*(sqrt(linspace(0,1,N+1)') + aggrTerm);
        %od = 4*(sqrt(linspace(0,1,N+1)') + 1/4);

        %ovrdFact = (ovrdSm - 1) / sqrt(ovrdPos/(N+1));
        %ovrdFact = ovrdSm / sqrt(echoBandRange(floor(length(echoBandRange)/2))/(N+1));
        %od = ovrdFact*sqrt(linspace(0,1,N+1))' + 1;
        %od = ovrdSm*ones(N+1,1).*abs(WFb(:,dIdx))/(max(abs(WFb(:,dIdx)))+1e-10);

        %od = ovrdSm*ones(N+1,1);
        %od = ovrdSm*WFbD.*(sqrt(linspace(0,1,N+1))' + 1);

        od = ovrdSm*(sqrt(linspace(0,1,N+1))' + 1);
        %od = 4*(sqrt(linspace(0,1,N+1))' + 1);

        %od = 2*ones(N+1,1);
        %od = 2*ones(N+1,1);
        %sshift = ((1-hnled)*2-1).^3+1;
        sshift = ones(N+1,1);

        hnled = hnled.^(od.*sshift);

        %if hnlg > 0.75
            %if (suppState ~= 0)
            % transCtr = 0;
            %end
        % suppState = 0;
        %elseif hnlg < 0.6 & hnlg > 0.2
        % suppState = 1;
        %elseif hnlg < 0.1
            %hnled = zeros(N+1, 1);
            %if (suppState ~= 2)
            % transCtr = 0;
            %end
        % suppState = 2;
        %else
        % if (suppState ~= 2)
        % transCtr = 0;
        % end
        % suppState = 2;
        %end
        %if suppState == 0
        % hnled = ones(N+1, 1);
        %elseif suppState == 2
        % hnled = zeros(N+1, 1);
        %end
        %hnled(find(hnled < 0.1)) = 0;
        %hnled = hnled.^2;
        %if transCtr < 5
            %hnl = 0.75*hnl + 0.25*hnled;
        % transCtr = transCtr + 1;
        %else
            hnl = hnled;
        %end
        %hnled(find(hnled < 0.05)) = 0;
        ef = ef.*(hnl);

        %ef = ef.*(min(1 - cohxd, cohed).^2);
        %ef = ef.*((1-cohxd).^2);

        ovrdV(kk) = ovrdSm;
        %ovrdV(kk) = dIdx;
        %ovrdV(kk) = divergeFact;
        %hnledAvg(kk) = 1-mean(1-cohedFast(echoBandRange));
        hnledAvg(kk) = 1-mean(1-cohed(echoBandRange));
        hnlxdAvg(kk) = 1-mean(cohxd(echoBandRange));
        %hnlxdAvg(kk) = cohxd(5);
        %hnlSortQV(kk) = mean(hnled);
        hnlSortQV(kk) = hnlPrefAvgLow;
        hnlPrefAvgV(kk) = hnlPrefAvg;
        %hnlAvg(kk) = propLow;
        %ef(N/2:end) = 0;
        %ner = (sum(Sd) ./ (sum(Se.*(hnl.^2)) + 1e-10));

        % Comfort noise
        if (CNon)
            snn=sqrt(Sym);
            snn(1)=0; % Reject LF noise
            Un=snn.*exp(j*2*pi.*[0;rand(N-1,1);0]);

            % Weight comfort noise by suppression
            Un = sqrt(1-hnled.^2).*Un;
            Fmix = ef + Un;
        else
            Fmix = ef;
        end

    % Overlap and add in time domain for smoothness
    tmp = [Fmix ; flipud(conj(Fmix(2:N)))];
    mixw = wins.*real(ifft(tmp));
    mola = mbuf(end-N+1:end) + mixw(1:N);
    mbuf = mixw;
    ercn(pos:pos+N-1) = mola;%%%%%-------------you can hear the effect by sound(10*ercn,16000),add by Shichaog
    end % NLPon

    % Filter update
    % Ek2 = Ek ./(12*pn + 0.001); % Normalized error
    %     Ek2 = Ek2 * divergeFact;
    Ek2 = Ek ./(pn + 0.001); % Normalized error
    %Ek2 = Ek ./(100*pn + 0.001); % Normalized error

    %divergeIdx = find(abs(Ek) > abs(DD));
    %divergeIdx = find(Se > Sd);
    %threshMod = threshold*ones(N+1,1);
    %if length(divergeIdx) > 0
    %if sum(abs(Ek)) > sum(abs(DD))
        %WFb(divergeIdx,:) = WFb(divergeIdx,:) .* repmat(sqrt(Sd(divergeIdx)./(Se(divergeIdx)+1e-10))),1,M);
        %Ek2(divergeIdx) = Ek2(divergeIdx) .* sqrt(Sd(divergeIdx)./(Se(divergeIdx)+1e-10));
        %Ek2(divergeIdx) = Ek2(divergeIdx) .* abs(DD(divergeIdx))./(abs(Ek(divergeIdx))+1e-10);
        %WFb(divergeIdx,:) = WFbOld(divergeIdx,:);
        %WFb = WFbOld;
        %threshMod(divergeIdx) = threshMod(divergeIdx) .* abs(DD(divergeIdx))./(abs(Ek(divergeIdx))+1e-10);
    % threshMod(divergeIdx) = threshMod(divergeIdx) .* sqrt(Sd(divergeIdx)./(Se(divergeIdx)+1e-10));
    %end

%absEf = max(abs(Ek2), threshold);
%absEf = ones(N+1,1)*threshold./absEf;
%absEf = max(abs(Ek2), threshMod);
%absEf = threshMod./absEf;
%Ek2 = Ek2.*absEf;

    %if sum(Se) <= sum(Sd)

    % mEk = mufb.*Ek2;
    % PP = conj(XFm).*(ones(M,1) * mEk')';
    % tmp = [PP ; flipud(conj(PP(2:N,:)))];
    % IFPP = real(ifft(tmp));
    % PH = IFPP(1:N,:);
    % tmp = fft([PH;zeros(N,M)]);
    % FPH = tmp(1:N+1,:);
    % %WFbOld = WFb;
    % WFb = WFb + FPH;

    %else
    % WF = WFbOld;
    %end

% Shift old FFTs
    XFm(:,2:end) = XFm(:,1:end-1);
    YFm(:,2:end) = YFm(:,1:end-1);
    xfwm(:,2:end) = xfwm(:,1:end-1);
    dfm(:,2:end) = dfm(:,1:end-1);

%if mod(kk, floor(Nb/50)) == 0
    % fprintf(1, '.');
%end

if mod(kk, floor(Nb/100)) == 0
%if mod(kk, floor(Nb/500)) == 0
        %progressbar(kk/Nb);
        %figure(5)
        %plot(abs(WFb));
        %legend('1','2','3','4','5','6','7','8','9','10','11','12');
        %title(kk*N/fs);
        %figure(6)
        %plot(WFbD);
        %figure(6)
        %plot(threshMod)
        %if length(divergeIdx) > 0
        % plot(abs(DD))
        % hold on
        % plot(abs(Ek), 'r')
        % hold off
            %plot(min(sqrt(Sd./(Se+1e-10)),1))
            %axis([0 N 0 1]);
        %end
        %figure(6)
        %plot(cohedFast);
        %axis([1 N+1 0 1]);
        %plot(WFbEn);

        %figure(7)
        %plot(weight);
        %plot([cohxd 1-cohed]);
        %plot([cohxd 1-cohed 1-cohedFast hnled]);
        %plot([cohxd cohxdFast/max(cohxdFast)]);
        %legend('cohxd', '1-cohed', '1-cohedFast');
        %axis([1 65 0 1]);
        %pause(0.5);
        %overdrive
    end
end
%progressbar(1);

%figure(2);
%plot([feat(:,1) feat(:,2)+1 feat(:,3)+2 mfeat+3]);
%plot([feat(:,1) mfeat+1]);

%figure(3);
%plot(10*log10([dri erifb erifb3 ericn]));
%legend('Near-end','Error','Post NLP','Final',4);
% Compensate for delay
%ercn=[ercn(N+1:end);zeros(N,1)];
%ercn_=[ercn_(N+1:end);zeros(N,1)];

%figure(11);
%plot(cohxdSlow);

%figure(12);
%surf(cohxdSlow);
%shading interp;

%figure(13);
%plot(overdriveM);

%figure(14);
%surf(overdriveM);
%shading interp;

figure(10);
t = (0:Nb)*N/fs;
rrinSubSamp = rrin(N*(1:(Nb+1)));
plot(t, rrinSubSamp/max(abs(rrinSubSamp)),'b');
hold on
plot(t, hnledAvg, 'r');
plot(t, hnlxdAvg, 'g');
plot(t, hnlSortQV, 'y');
plot(t, hnlLocalMinV, 'k');
plot(t, cohxdLocalMinV, 'c');
plot(t, hnlPrefAvgV, 'm');
%plot(t, cohxdAvg, 'r');
%plot(cohxdFastAvg, 'r');
%plot(cohxdAvgBad, 'k');
%plot(t, cohedAvg, 'k');
%plot(t, 1-cohedFastAvg, 'k');
%plot(ssin(N*(1:floor(length(ssin)/N)))/max(abs(ssin)));
%plot(echoBands,'r');
%plot(overdrive, 'g');
%plot(erfb(N*(1:floor(length(erfb)/N)))/max(abs(erfb)));
hold off
%tight x;

% figure(11)
% plot(t, ovrdV);
%tightx;
%plot(mfeat,'r');
%plot(1-cohxyp_,'r');
%plot(Hnlxydp,'y');
%plot(hnledp,'k');
%plot(Hnlxydp, 'c');
%plot(ccohpd_,'k');
%plot(supplot_, 'g');
%plot(ones(length(mfeat),1)*rr1_, 'k');
%plot(ones(length(mfeat),1)*rr2_, 'k');
%plot(N*(1:length(feat)), feat);
%plot(Sep_,'r');
%axis([1 floor(length(erfb)/N) -1 1])
%hold off
%plot(10*log10([Se_, Sx_, Seu_, real(sf_.*conj(sf_))]));
%legend('Se','Sx','Seu','S');
%figure(5)
%plot([ercn ercn_]);

% figure(12)
% plot(t, dIdxV);
%plot(t, SLxV);
%tightx;

%figure(13)
%plot(t, [ekEnV dkEnV]);
%plot(t, dkEnV./(ekEnV+1e-10));
%tightx;

%close(hh);
%spclab(fs,ssin,erfb,ercn,'outxd.pcm');
%spclab(fs,rrin,ssin,erfb,1.78*ercn,'vqeOut-1.pcm');
%spclab(fs,erfb,'aecOutLp.pcm');
%spclab(fs,rrin,ssin,erfb,1.78*ercn,'aecOut25.pcm','vqeOut-1.pcm');
%spclab(fs,rrin,ssin,erfb,ercn,'aecOut-mba.pcm');
%spclab(fs,rrin,ssin,erfb,ercn,'aecOut.pcm');
%spclab(fs, ssin, erfb, ercn, 'out0.pcm');

speex AEC算法

和WebRTC一樣也是採用頻域分塊自適應濾波方法，不同的是權重調整的方式變化，我這邊測試效果是計算量比WebRTC的大，且效果調節的沒有WebRTC的好。這裏也給出speex的源代碼和測試方法。

%    Copyright (C) 2012      Waves Audio LTD
%    Copyright (C) 2003-2008 Jean-Marc Valin
%
%    File: speex_mdf.m
%    Echo canceller based on the MDF algorithm (see below)
% 
%    Redistribution and use in source and binary forms, with or without
%    modification, are permitted provided that the following conditions are
%    met:
% 
%    1. Redistributions of source code must retain the above copyright notice,
%    this list of conditions and the following disclaimer.
% 
%    2. Redistributions in binary form must reproduce the above copyright
%    notice, this list of conditions and the following disclaimer in the
%    documentation and/or other materials provided with the distribution.
% 
%    3. The name of the author may not be used to endorse or promote products
%    derived from this software without specific prior written permission.
% 
%    THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
%    IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
%    OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
%    DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
%    INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
%    (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
%    SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
%    HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
%    STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
%    ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
%    POSSIBILITY OF SUCH DAMAGE.
%
%    Notes from original mdf.c:
%
%    The echo canceller is based on the MDF algorithm described in:
% 
%    J. S. Soo, K. K. Pang Multidelay block frequency adaptive filter, 
%    IEEE Trans. Acoust. Speech Signal Process., Vol. ASSP-38, No. 2, 
%    February 1990.
%    
%    We use the Alternatively Updated MDF (AUMDF) variant. Robustness to 
%    double-talk is achieved using a variable learning rate as described in:
%    
%    Valin, J.-M., On Adjusting the Learning Rate in Frequency Domain Echo 
%    Cancellation With Double-Talk. IEEE Transactions on Audio,
%    Speech and Language Processing, Vol. 15, No. 3, pp. 1030-1034, 2007.
%    http://people.xiph.org/~jm/papers/valin_taslp2006.pdf
%    
%    There is no explicit double-talk detection, but a continuous variation
%    in the learning rate based on residual echo, double-talk and background
%    noise.
%    
%    Another kludge that seems to work good: when performing the weight
%    update, we only move half the way toward the "goal" this seems to
%    reduce the effect of quantization noise in the update phase. This
%    can be seen as applying a gradient descent on a "soft constraint"
%    instead of having a hard constraint.
%    
%    Notes for this file:
%
%    Usage: 
%
%       speex_mdf_out = speex_mdf(Fs, u, d, filter_length, frame_size, dbg_var_name);
%       
%       Fs                  sample rate
%       u                   speaker signal, column vector in range [-1; 1]
%       d                   microphone signal, column vector in range [-1; 1]
%       filter_length       typically 250ms, i.e. 4096 @ 16k FS 
%                           must be a power of 2
%       frame_size          typically 8ms, i.e. 128 @ 16k Fs 
%                           must be a power of 2
%       dbg_var_name        internal state variable name to trace. 
%                           Default: 'st.leak_estimate'.
%
%    Jonathan Rouach <[email protected]>
%    

function  speex_mdf_out = speex_mdf(Fs, u, d, filter_length, frame_size, dbg_var_name)

fprintf('Starting Speex MDF (PBFDAF) algorithm.\n');

st = speex_echo_state_init_mc_mdf(frame_size, filter_length, 1, 1, Fs);

% which variable to trace
if nargin<6
    dbg_var_name = 'st.leak_estimate';
end
dbg = init_dbg(st, length(u));

[e, dbg] = main_loop(st, float_to_short(u), float_to_short(d), dbg);

speex_mdf_out.e = e/32768.0;
speex_mdf_out.var1 = dbg.var1;

    function x = float_to_short(x)
        x = x*32768.0;
        x(x< -32767.5) = -32768;
        x(x>  32766.5) =  32767;
        x = floor(0.5+x);
    end

    function [e, dbg] = main_loop(st, u, d, dbg)

        e = zeros(size(u));
        y = zeros(size(u));

        % prepare waitbar
        try h_wb = waitbar(0, 'Processing...'); catch; end
        end_point = length(u);

        for n = 1:st.frame_size:end_point
            nStep = floor(n/st.frame_size)+1;

            if mod(nStep, 128)==0 && update_waitbar_check_wasclosed(h_wb, n, end_point, st.sampling_rate)
                break;
            end

            u_frame = u(n:n+st.frame_size-1);
            d_frame = d(n:n+st.frame_size-1);

            [out, st] = speex_echo_cancellation_mdf(st, d_frame, u_frame);

            e(n:n+st.frame_size-1) = out*2;
            y(n:n+st.frame_size-1) = d_frame - out;
            dbg.var1(:, nStep) = reshape( eval(dbg_var_name),  numel(eval(dbg_var_name)), 1);

        end

        try close(h_wb); catch; end

    end
    function st = speex_echo_state_init_mc_mdf(frame_size, filter_length, nb_mic, nb_speakers, sample_rate)

        st.K = nb_speakers;
        st.C = nb_mic;
        C=st.C;
        K=st.K;

        st.frame_size = frame_size;
        st.window_size = 2*frame_size;
        N = st.window_size;
        st.M = fix((filter_length+st.frame_size-1)/frame_size);
        M = st.M;
        st.cancel_count=0;
        st.sum_adapt = 0;
        st.saturated = 0;
        st.screwed_up = 0;

        %    /* This is the default sampling rate */
        st.sampling_rate = sample_rate;
        st.spec_average = (st.frame_size)/( st.sampling_rate);
        st.beta0 = (2.0*st.frame_size)/st.sampling_rate;
        st.beta_max = (.5*st.frame_size)/st.sampling_rate;
        st.leak_estimate = 0;

        st.e = zeros(N, C);
        st.x = zeros(N, K);
        st.input = zeros(st.frame_size, C);
        st.y = zeros(N, C);
        st.last_y = zeros(N, C);
        st.Yf = zeros(st.frame_size+1, 1);
        st.Rf = zeros(st.frame_size+1, 1);
        st.Xf = zeros(st.frame_size+1, 1);
        st.Yh = zeros(st.frame_size+1, 1);
        st.Eh = zeros(st.frame_size+1, 1);

        st.X = zeros(N, K, M+1);
        st.Y = zeros(N, C);
        st.E = zeros(N, C);
        st.W = zeros(N, K, M, C);
        st.foreground = zeros(N, K, M, C);
        st.PHI = zeros(frame_size+1, 1);
        st.power = zeros(frame_size+1, 1);
        st.power_1 = ones((frame_size+1), 1);
        st.window = zeros(N, 1);
        st.prop = zeros(M, 1);
        st.wtmp = zeros(N, 1);

        st.window = .5-.5*cos(2*pi*((1:N)'-1)/N);

        % /* Ratio of ~10 between adaptation rate of first and last block */
        decay = exp(-1/M);
        st.prop(1, 1) = .7;
        for i=2:M
            st.prop(i, 1) = st.prop(i-1, 1) * decay;
        end

        st.prop = (.8 * st.prop)./sum(st.prop);

        st.memX = zeros(K, 1);
        st.memD = zeros(C, 1);
        st.memE = zeros(C, 1);
        st.preemph = .98;
        if (st.sampling_rate<12000)
            st.notch_radius = .9;
        elseif (st.sampling_rate<24000)
            st.notch_radius = .982;
        else
            st.notch_radius = .992;
        end

        st.notch_mem = zeros(2*C, 1);
        st.adapted = 0;
        st.Pey = 1;
        st.Pyy = 1;

        st.Davg1 = 0; st.Davg2 = 0;
        st.Dvar1 = 0; st.Dvar2 = 0;
    end

    function dbg = init_dbg(st, len)
        dbg.var1 = zeros(numel(eval(dbg_var_name)), fix(len/st.frame_size));
    end

    function [out, st] = speex_echo_cancellation_mdf(st, in, far_end)

        N = st.window_size;
        M = st.M;
        C = st.C;
        K = st.K;

        Pey_cur = 1;
        Pyy_cur = 1;

        out = zeros(st.frame_size, C);

        st.cancel_count = st.cancel_count + 1;

        %ss=.35/M;
        ss = 0.5/M;
        ss_1 = 1-ss;

        for chan = 1:C
            % Apply a notch filter to make sure DC doesn't end up causing problems
            [st.input(:, chan), st.notch_mem(:, chan)] = filter_dc_notch16(in(:, chan), st.notch_radius, st.frame_size, st.notch_mem(:, chan));
            % Copy input data to buffer and apply pre-emphasis
            for i=1:st.frame_size
                tmp32 = st.input(i, chan)- (st.preemph* st.memD(chan));
                st.memD(chan) = st.input(i, chan);
                st.input(i, chan) = tmp32;
            end
        end

        for speak = 1:K
            for i =1:st.frame_size
                st.x(i, speak) = st.x(i+st.frame_size, speak);
                tmp32 = far_end(i, speak) - st.preemph * st.memX(speak);
                st.x(i+st.frame_size, speak) = tmp32;
                st.memX(speak) = far_end(i, speak);
            end
        end

        % Shift memory
        st.X = circshift(st.X, [0, 0, 1]);

        for speak = 1:K
            %  Convert x (echo input) to frequency domain
            % MATLAB_MATCH: we divide by N to get values as in speex
            st.X(:, speak, 1) = fft(st.x(:, speak)) /N;
        end

        Sxx = 0;
        for speak = 1:K
            Sxx = Sxx + sum(st.x(st.frame_size+1:end, speak).^2);
            st.Xf = abs(st.X(1:st.frame_size+1, speak, 1)).^2;
        end

        Sff = 0;
        for chan = 1:C

            %  Compute foreground filter
            st.Y(:, chan) = 0;
            for speak=1:K
                for j=1:M
                    st.Y(:, chan) = st.Y(:, chan) + st.X(:, speak, j) .* st.foreground(:, speak, j, chan);
                end
            end
            % MATLAB_MATCH: we multiply by N to get values as in speex
            st.e(:, chan) = ifft(st.Y(:, chan)) * N;
            st.e(1:st.frame_size, chan) = st.input(:, chan) - st.e(st.frame_size+1:end, chan);
            % st.e : [out foreground | leak foreground ]
            Sff = Sff + sum(abs(st.e(1:st.frame_size, chan)).^2);

        end

        % Adjust proportional adaption rate */
        if (st.adapted)
            st.prop = mdf_adjust_prop (st.W, N, M, C, K);
        end

        % Compute weight gradient */
        if (st.saturated == 0)
            for chan = 1:C
                for speak = 1:K
                    for j=M:-1:1
                        st.PHI = [st.power_1; st.power_1(end-1:-1:2)] .* st.prop(j) .* conj(st.X(:, speak, (j+1))) .* st.E(:, chan);
                        st.W(:, j) = st.W(:, j) + st.PHI;
                    end
                end
            end
        else
            st.saturated = st.saturated -1;
        end

        %FIXME: MC conversion required */
        % Update weight to prevent circular convolution (MDF / AUMDF)
        for chan = 1:C
            for speak = 1:K
                for j = 1:M
                    % This is a variant of the Alternatively Updated MDF (AUMDF) */
                    % Remove the "if" to make this an MDF filter */
                    if (j==1 || mod(2+st.cancel_count,(M-1)) == j)
                        st.wtmp = ifft(st.W(:, speak, j, chan));
                        st.wtmp(st.frame_size+1:N) = 0;
                        st.W(:, speak, j, chan) = fft(st.wtmp);
                    end
                end
            end
        end

        % So we can use power_spectrum_accum */
        st.Yf = zeros(st.frame_size+1, 1);
        st.Rf = zeros(st.frame_size+1, 1);
        st.Xf = zeros(st.frame_size+1, 1);

        Dbf = 0;

        for chan = 1:C
            st.Y(:, chan) = 0;
            for speak=1:K
                for j=1:M
                    st.Y(:, chan) = st.Y(:, chan) + st.X(:, speak, j) .* st.W(:, speak, j, chan);
                end
            end
            % MATLAB_MATCH: we multiply by N to get values as in speex
            st.y(:,chan) = ifft(st.Y(:,chan)) * N;
            % st.y : [ ~ | leak background ]
        end

        See = 0;

        % Difference in response, this is used to estimate the variance of our residual power estimate */
        for chan = 1:C
            st.e(1:st.frame_size, chan) = st.e(st.frame_size+1:N, chan) - st.y(st.frame_size+1:N, chan);
            Dbf = Dbf + 10 + sum(abs(st.e(1:st.frame_size, chan)).^2);
            st.e(1:st.frame_size, chan) = st.input(:, chan) - st.y(st.frame_size+1:N, chan);
            % st.e : [ out background | leak foreground ]
           See = See + sum(abs(st.e(1:st.frame_size, chan)).^2);
        end

        % Logic for updating the foreground filter */

        % For two time windows, compute the mean of the energy difference, as well as the variance */
        VAR1_UPDATE = .5;
        VAR2_UPDATE = .25;
        VAR_BACKTRACK = 4;
        MIN_LEAK = .005;

        st.Davg1 = .6*st.Davg1 + .4*(Sff-See);
        st.Davg2 = .85*st.Davg2 + .15*(Sff-See);
        st.Dvar1 = .36*st.Dvar1 + .16*Sff*Dbf;
        st.Dvar2 = .7225*st.Dvar2 + .0225*Sff*Dbf;

        update_foreground = 0;

        % Check if we have a statistically significant reduction in the residual echo */
        % Note that this is *not* Gaussian, so we need to be careful about the longer tail */
        if (Sff-See)*abs(Sff-See) > (Sff*Dbf)
            update_foreground = 1;
        elseif (st.Davg1* abs(st.Davg1) > (VAR1_UPDATE*st.Dvar1))
            update_foreground = 1;
        elseif (st.Davg2* abs(st.Davg2) > (VAR2_UPDATE*(st.Dvar2)))
            update_foreground = 1;
        end

        % Do we update? */
        if (update_foreground)

            st.Davg1 = 0;
            st.Davg2 = 0;
            st.Dvar1 = 0;
            st.Dvar2 = 0;
            st.foreground = st.W;
            % Apply a smooth transition so as to not introduce blocking artifacts */
            for chan = 1:C
                st.e(st.frame_size+1:N, chan) = (st.window(st.frame_size+1:N) .* st.e(st.frame_size+1:N, chan)) + (st.window(1:st.frame_size) .* st.y(st.frame_size+1:N, chan));
            end
        else
            reset_background=0;
            % Otherwise, check if the background filter is significantly worse */

            if (-(Sff-See)*abs(Sff-See)> VAR_BACKTRACK*(Sff*Dbf))
                reset_background = 1;
            end
            if ((-st.Davg1 * abs(st.Davg1))> (VAR_BACKTRACK*st.Dvar1))
                reset_background = 1;
            end
            if ((-st.Davg2* abs(st.Davg2))> (VAR_BACKTRACK*st.Dvar2))
                reset_background = 1;
            end

            if (reset_background)

                % Copy foreground filter to background filter */
                st.W = st.foreground;

                % We also need to copy the output so as to get correct adaptation */
                for chan = 1:C
                    st.y(st.frame_size+1:N, chan) = st.e(st.frame_size+1:N, chan);
                    st.e(1:st.frame_size, chan) = st.input(:, chan) - st.y(st.frame_size+1:N, chan);
                end

                See = Sff;
                st.Davg1 = 0;
                st.Davg2 = 0;
                st.Dvar1 = 0;
                st.Dvar2 = 0;
            end
        end

        Sey = 0;
        Syy = 0;
        Sdd = 0;

        for chan = 1:C

            % Compute error signal (for the output with de-emphasis) */
            for i=1:st.frame_size
                tmp_out = st.input(i, chan)- st.e(i+st.frame_size, chan);
                tmp_out = tmp_out + st.preemph * st.memE(chan);
                %  This is an arbitrary test for saturation in the microphone signal */
                if (in(i,chan) <= -32000 || in(i,chan) >= 32000)
                    if (st.saturated == 0)
                        st.saturated = 1;
                    end
                end
                out(i, chan) = tmp_out;
                st.memE(chan) = tmp_out;
            end

            % Compute error signal (filter update version) */
            st.e(st.frame_size+1:N, chan) = st.e(1:st.frame_size, chan);
            st.e(1:st.frame_size, chan) = 0;
            % st.e : [ zeros | out background ]

            % Compute a bunch of correlations */
            % FIXME: bad merge */
            Sey = Sey + sum(st.e(st.frame_size+1:N, chan) .* st.y(st.frame_size+1:N, chan));
            Syy = Syy + sum(st.y(st.frame_size+1:N, chan).^2);
            Sdd = Sdd + sum(st.input.^2);

            % Convert error to frequency domain */
            % MATLAB_MATCH: we divide by N to get values as in speex
            st.E = fft(st.e) / N;

            st.y(1:st.frame_size, chan) = 0;
            % MATLAB_MATCH: we divide by N to get values as in speex
            st.Y = fft(st.y) / N;

            % Compute power spectrum of echo (X), error (E) and filter response (Y) */
            st.Rf = abs(st.E(1:st.frame_size+1,chan)).^2;
            st.Yf = abs(st.Y(1:st.frame_size+1,chan)).^2;
        end

        % Do some sanity check */
        if (~(Syy>=0 && Sxx>=0 && See >= 0))
            % Things have gone really bad */
            st.screwed_up = st.screwed_up + 50;
            out = out*0;
        elseif Sff > Sdd+ N*10000
            % AEC seems to add lots of echo instead of removing it, let's see if it will improve */
            st.screwed_up = st.screwed_up + 1;
        else
            % Everything's fine */
            st.screwed_up=0;
        end

        if (st.screwed_up>=50)
            disp('Screwed up, full reset');
            st = speex_echo_state_reset_mdf(st);
        end

        % Add a small noise floor to make sure not to have problems when dividing */
        See = max(See, N* 100);

        for speak = 1:K
            Sxx = Sxx + sum(st.x(st.frame_size+1:end, speak).^2);
            st.Xf = abs(st.X(1:st.frame_size+1, speak, 1)).^2;
        end

        % Smooth far end energy estimate over time */
        st.power = ss_1*st.power+ 1 + ss*st.Xf;

        % Compute filtered spectra and (cross-)correlations */

        Eh_cur = st.Rf - st.Eh;
        Yh_cur = st.Yf - st.Yh;
        Pey_cur = Pey_cur + sum(Eh_cur.*Yh_cur) ;
        Pyy_cur = Pyy_cur + sum(Yh_cur.^2);
        st.Eh = (1-st.spec_average)*st.Eh + st.spec_average*st.Rf;
        st.Yh = (1-st.spec_average)*st.Yh + st.spec_average*st.Yf;

        Pyy = sqrt(Pyy_cur);
        Pey = Pey_cur/Pyy;

        % Compute correlation updatete rate */
        tmp32 = st.beta0*Syy;
        if (tmp32 > st.beta_max*See)
            tmp32 = st.beta_max*See;
        end
        alpha = tmp32/ See;
        alpha_1 = 1- alpha;

        % Update correlations (recursive average) */
        st.Pey = alpha_1*st.Pey + alpha*Pey;
        st.Pyy = alpha_1*st.Pyy + alpha*Pyy;

        if st.Pyy<1
            st.Pyy =1;
        end

        % We don't really hope to get better than 33 dB (MIN_LEAK-3dB) attenuation anyway */
        if st.Pey< MIN_LEAK * st.Pyy
            st.Pey = MIN_LEAK * st.Pyy;
        end

        if (st.Pey> st.Pyy)
            st.Pey = st.Pyy;
        end

        % leak_estimate is the linear regression result */
        st.leak_estimate = st.Pey/st.Pyy;

        % This looks like a stupid bug, but it's right (because we convert from Q14 to Q15) */
        if (st.leak_estimate > 16383)
            st.leak_estimate = 32767;
        end

        % Compute Residual to Error Ratio */
        RER = (.0001*Sxx + 3.*st.leak_estimate*Syy) / See;
        % Check for y in e (lower bound on RER) */
        if (RER < Sey*Sey/(1+See*Syy))
            RER = Sey*Sey/(1+See*Syy);
        end
        if (RER > .5)
            RER = .5;
        end

        % We consider that the filter has had minimal adaptation if the following is true*/
        if (~st.adapted && st.sum_adapt > M && st.leak_estimate*Syy > .03*Syy)
            st.adapted = 1;
        end

        if (st.adapted)
            % Normal learning rate calculation once we're past the minimal adaptation phase */
            for i=1:st.frame_size+1

                % Compute frequency-domain adaptation mask */
                r = st.leak_estimate*st.Yf(i);
                e = st.Rf(i)+1;
                if (r>.5*e)
                    r = .5*e;
                end
                r = 0.7*r + 0.3*(RER*e);
                %st.power_1[i] = adapt_rate*r/(e*(1+st.power[i]));*/
                st.power_1(i) = (r/(e*st.power(i)+10));
            end
        else
            % Temporary adaption rate if filter is not yet adapted enough */
            adapt_rate=0;

            if (Sxx > N* 1000)

                tmp32 = 0.25* Sxx;
                if (tmp32 > .25*See)
                    tmp32 = .25*See;
                end
                adapt_rate = tmp32/ See;
            end
            st.power_1 = adapt_rate./(st.power+10);


            % How much have we adapted so far? */
            st.sum_adapt = st.sum_adapt+adapt_rate;
        end

        % FIXME: MC conversion required */
        st.last_y(1:st.frame_size) = st.last_y(st.frame_size+1:N);
        if (st.adapted)
            % If the filter is adapted, take the filtered echo */
            st.last_y(st.frame_size+1:N) = in-out;
        end

    end

    function [out,mem] = filter_dc_notch16(in, radius, len, mem)
        out = zeros(size(in));
        den2 = radius*radius + .7*(1-radius)*(1-radius);
        for i=1:len
            vin = in(i);
            vout = mem(1) + vin;
            mem(1) = mem(2) + 2*(-vin + radius*vout);
            mem(2) = vin - (den2*vout);
            out(i) = radius*vout; 
        end

    end

    function prop = mdf_adjust_prop(W, N, M, C, K)
        prop = zeros(M,1);
        for i=1:M
            tmp = 1;
            for chan=1:C
                for speak=1:K
                    tmp = tmp + sum(abs(W(1:N/2+1, K, i, C)).^2);
                end
            end
            prop(i) = sqrt(tmp);
        end
        max_sum = max(prop, 1);
        prop = prop + .1*max_sum;
        prop_sum = 1+sum(prop);
        prop = .99*prop / prop_sum;
    end

    % Resets echo canceller state */
    function st = speex_echo_state_reset_mdf(st)

        st.cancel_count=0;
        st.screwed_up = 0;
        N = st.window_size;
        M = st.M;
        C=st.C;
        K=st.K;

        st.e = zeros(N, C);
        st.x = zeros(N, K);
        st.input = zeros(st.frame_size, C);
        st.y = zeros(N, C);
        st.last_y = zeros(N, C);
        st.Yf = zeros(st.frame_size+1, 1);
        st.Rf = zeros(st.frame_size+1, 1);
        st.Xf = zeros(st.frame_size+1, 1);
        st.Yh = zeros(st.frame_size+1, 1);
        st.Eh = zeros(st.frame_size+1, 1);

        st.X = zeros(N, K, M+1);
        st.Y = zeros(N, C);
        st.E = zeros(N, C);
        st.W = zeros(N, K, M, C);
        st.foreground = zeros(N, K, M, C);
        st.PHI = zeros(N, 1);
        st.power = zeros(st.frame_size+1, 1);
        st.power_1 = ones((st.frame_size+1), 1);
        st.window = zeros(N, 1);
        st.prop = zeros(M, 1);
        st.wtmp = zeros(N, 1);

        st.memX = zeros(K, 1);
        st.memD = zeros(C, 1);
        st.memE = zeros(C, 1);

        st.saturated = 0;
        st.adapted = 0;
        st.sum_adapt = 0;
        st.Pey = 1;
        st.Pyy = 1;
        st.Davg1 = 0;
        st.Davg2 = 0;
        st.Dvar1 = 0;
        st.Dvar2 = 0;


    end

    function was_closed = update_waitbar_check_wasclosed(h, n, end_point, Fs)
        was_closed = 0;

        % update waitbar
        try
            waitbar(n/end_point, h, ['Processing... ', num2str(n/Fs, '%.2f'), 's / ', num2str(end_point/Fs, '%.2f'), 's' ]);
        catch ME
            % if it's no longer there (closed by user)
            if (strcmp(ME.identifier(1:length('MATLAB:waitbar:')), 'MATLAB:waitbar:'))
                was_closed = 1; % then get out of the loop
            end
        end

    end

end

測試方法

首先需要自己讀取文件並設置相關的初始值
給出自己的一個樣例

fid=fopen('near.pcm', 'rb'); % Load far end
ssin=fread(fid,inf,'float32');
fid=fopen('far.pcm', 'rb'); % Load fnear end
rrin=fread(fid,inf,'float32');
ssin=ssin(1:4096*200);
rrin=rrin(1:4096*200);
Fs=16000;
filter_length=4096;
frame_size=128;
speex_mdf_out = speex_mdf(Fs, rrin, ssin, filter_length, frame_size);

執行完之後，需要播放出來聽：

sound(speex_mdf_out.e,16000)

代碼里名詞術語

RERL:ERL+ERLE
RERL:residual_echo_return_loss
ERL:echo_return_loss
ERLE:echo_return_loss_enhancement
psd:power spectral density 功率譜密度
x: far end
d: near end
e: error
s: psd
nlp:non-linear processing

轉載_語音自適應回聲消除（AEC）算法

自適應回聲消除算法

回聲消除原理

NLMS權重調整

功率計算

濾波

誤差估計

自適應調節

功率譜密度和相關性計算

NLP

發散處理

平滑濾波器係數和抑制指數

輸出回聲消除後的信號

speex AEC算法

測試方法

代碼里名詞術語

apisix~helm方式的部署到k8s

firmeye - IoT固件漏洞挖掘工具

轉載_[原創] 使用Tracealyzer分析RTOS需掌握的基本概念

轉載_怎樣使用Yocto項目爲樹莓派構建GNU / Linux發行版

轉載_樹莓派4B的詳細資料說明

轉載_樹莓派4B下的usart串口測試

轉載_Git分支管理策略

https://yachay.unat.edu.pe/blog/index.php?comment_area=format_blog&comment_component=blog&comment_co

linux以太網驅動總結