[Speech denoising] Based on matlab spectral subtraction voice denoising [including Matlab source code 571]

1. Introduction

The most commonly used method in speech denoising is spectral subtraction. Spectral subtraction is an earlier developed and more mature speech denoising algorithm. This algorithm uses the characteristic that additive noise is not related to speech, and assumes that noise is statistically stable. Under the premise of, the estimated value of the noise spectrum measured without the speech gap is used to replace the spectrum with the noise during the speech, and subtract it from the noisy speech spectrum to obtain the estimated value of the speech spectrum. Spectral subtraction has the characteristics of simple algorithm and small amount of calculation, which is convenient for fast processing, and can often obtain a higher output signal-to-noise ratio, so it is widely used. The disadvantage of the classic form of this algorithm is that it will produce "musical noise" with certain rhythmic fluctuations and sounds similar to music after processing.

After converting to the frequency domain, these peaks sound like multi-frequency tones with random changes in frequency from frame to frame. This situation is especially obvious in the unvoiced segment. This kind of "noise" caused by half-wave rectification is called "music" noise". Fundamentally, the main reasons that usually cause music noise are:
(1) Non-linear processing of the negative part of the spectral subtraction algorithm
(2) Inaccurate estimation of the noise spectrum
(3) The suppression function (gain function) is relatively high. Large variability
1 principle
Insert picture description here

Second, the source code

clear all; clc; close all;

[xx, fs] = wavread('C5_2_y.wav');           % 读入数据文件
xx=xx-mean(xx);                         % 消除直流分量
x=xx/max(abs(xx));                      % 幅值归一化

IS=0.25;                                % 设置前导无话段长度
wlen=200;                               % 设置帧长为25ms
inc=80;                                 % 设置帧移为10ms
SNR=5;                                  % 设置信噪比SNR
N=length(x);                            % 信号长度
time=(0:N-1)/fs;                        % 设置时间

signal=awgn(x,SNR,'measured','db');               % 叠加噪声
snr1=SNR_Calc(x,signal);            % 计算初始信噪比
NIS=fix((IS*fs-wlen)/inc +1);           % 求前导无话段帧数

a=4; b=0.001;                           % 设置参数a和b
output=SpectralSub(signal,wlen,inc,NIS,a,b);% 谱减
snr2=SNR_Calc(x,output);            % 计算谱减后的信噪比
snr=snr2-snr1;
fprintf('snr1=%5.4f   snr2=%5.4f   snr=%5.4f\n',snr1,snr2,snr);
unction frameout=enframe(x,win,inc)

nx=length(x(:));            % 取数据长度
nwin=length(win);           % 取窗长
if (nwin == 1)              % 判断窗长是否为1，若为1，即表示没有设窗函数
   len = win;               % 是，帧长=win
else
   len = nwin;              % 否，帧长=窗长
end
if (nargin < 3)             % 如果只有两个参数，设帧inc=帧长
   inc = len;
end
nf = fix((nx-len+inc)/inc); % 计算帧数
frameout=zeros(nf,len);            % 初始化
indf= inc*(0:(nf-1)).';     % 设置每帧在x中的位移量位置
inds = (1:len);             % 每帧数据对应1:len
frameout(:) = x(indf(:,ones(1,len))+inds(ones(nf,1),:));   % 对数据分帧
if (nwin > 1)               % 若参数中包括窗函数，把每帧乘以窗函数
    w = win(:)';            % 把win转成行数据
    function frameout=filpframe(x,win,inc)

[nf,len]=size(x);
nx=(nf-1) *inc+len;                 %原信号长度
frameout=zeros(nx,1);
nwin=length(win);                   % 取窗长
if (nwin ~= 1)                           % 判断窗长是否为1，若为1，即表示没有设窗函数
    winx=repmat(win',nf,1);
    x=x./winx;                          % 除去加窗的影响
    x(find(isinf(x)))=0;                %去除除0得到的Inf
end

Three, running results

Insert picture description here

Four, remarks