语音降噪_标准谱减法（附Python源码） @Learning Speech enhancement__1

 最近在看speech enhancement 内容，看完谱减法部分后，在网上找相应的代码来看，然后将MATLAB代码转成Python代码，顺便学习一下Python的使用。

谱减法的基础实现：

 

论文《Enhancement of speech corrupted by acoustic noise》提出的实现：

算法流程如下：

效果如下：

这是一段火车站附近的录音，噪声比较平稳；设置的VAD阈值是3dB，一般应用上设置的是6dB

虽然floor值的存在填充了频谱中一定的 “波谷”， 但可以看到还是有一些音乐噪声的存在。

 

 

代码段

Python代码如下：

1. #!/usr/bin/env python
   # encoding: utf-8
   '''
   @author: Kolor
   @license: 
   @contact: [email protected]
   @software: pycharm
   @file: spec_sub.py
   @time: 2020/12/21 23:44
   @desc:
     Implements the basic power spectral subtraction algorithm [1].
   
     Usage:  specsub(noisyFile, outputFile)
   
            noisyFile - noisy speech file in .wav format
            outputFile - enhanced output file in .wav format
   
      Algorithm uses the first 5 frames to estimate the noise psd, and
      then uses a very simple VAD algorithm for updating the noise psd.
      Alternatively, the VAD in the file ss_rdc.m (this folder) can be used.
   
     References:
      [1] Berouti, M., Schwartz, M., and Makhoul, J. (1979). Enhancement of speech
          corrupted by acoustic noise. Proc. IEEE Int. Conf. Acoust., Speech,
          Signal Processing, 208-211.
   
    Author: Philipos C. Loizou
   
    Copyright (c) 2006 by Philipos C. Loizou
    $Revision: 0.0 $  $Date: 10/09/2006 $
   -------------------------------------------------------------------------
   '''
   from typing import Type
   
   import librosa
   import numpy as np
   import sys
   import matplotlib.pyplot as plt
   import soundfile as sf
   
   
   def rh_spec_sub(infile, outfile):
       print(np.__version__)
       x, samp_rate = librosa.load(infile, sr=8000, mono=True)
       # sys.exit("pause")
       # =============== Initialize variables ===============
   
       # Frame size in samples
       frame_len = int(np.floor(20 * samp_rate / 1000))
       if np.remainder(frame_len, 2) == 1:
           frame_len = frame_len + 1
   
       # window overlap in percent of frame size
       overlap_percent = 50
       len1 = int(np.floor(frame_len * overlap_percent / 100))
       len2 = int(frame_len - len1)
   
       # VAD threshold in dB SNR
       vad_thresh = 3
       # power exponent
       alpha = 2.0
       FLOOR = 0.002
       G = 0.9
   
       # define window
       win = np.hanning(frame_len+1)[0:frame_len]
       win_gain = len2 / np.sum(win)  # normalization gain for overlap+add with 50% overlap
   
       # Noise magnitude calculations - assuming that the first 5 frames is noise/silence
   
       NFFT = int(2 * 2 ** next_pow2(frame_len))
       noise_mean = np.zeros(NFFT)
       j = 0
       for k in range(0, 5):
           windowed = np.multiply(win, x[j:j+frame_len])
           noise_mean = np.add(noise_mean, np.abs(np.fft.fft(windowed, NFFT)))
           j = j + frame_len
   
       noise_mu = noise_mean / 5
   
       # --- allocate memory and initialize various variables
       k = 0
       img = 1j  # np.sqrt(-1)
       x_old = np.zeros(len1)
       Nframs = int(np.floor(len(x)/len2) - 1)
       xfinal = np.zeros(Nframs * len2)
   
       # ===============================  Start Processing ==================================
       for n in range(0, Nframs):
           insign = np.multiply(win, x[k:k+frame_len])
           spec = np.fft.fft(insign, NFFT)
           sig = np.abs(spec)  # compute the magnitude
   
           # save the noisy phase information
           theta = np.angle(spec)
   
           SNR_seg = 10*np.log10(np.linalg.norm(sig, 2)**2 / np.linalg.norm(noise_mu, 2)**2)
   
           if alpha == 1.0:
               beta = berouti1(SNR_seg)
           else:
               beta = berouti(SNR_seg)
   
           # &&&&&&&&&&&&&&&
           sub_speech = sig ** alpha - beta * noise_mu ** alpha
           diffw = sub_speech - FLOOR * noise_mu ** alpha
   
           # floor negative component
           z = np.argwhere(diffw < 0)
           if len(z) is not 0:
               sub_speech[z] = FLOOR * noise_mu[z] ** alpha
   
           # --- implement a simple VAD detector - -------------
           if SNR_seg < vad_thresh:   # Update noise spectrum
               noise_temp = G * noise_mu ** alpha + (1 - G) * sig ** alpha
               noise_mu = noise_temp ** (1 / alpha)  # new noise spectrum
   
           # to ensure conjugate symmetry for real reconstruction
           sub_speech[int(NFFT/2) + 1: NFFT] = np.flipud(sub_speech[1: int(NFFT/2)])
   
           x_phase = (sub_speech**(1/alpha)) * (np.cos(theta) + img * (np.sin(theta)))
   
           # take the iFFT
           xi = np.real(np.fft.ifft(x_phase))
           # plt.plot(xi)
           # plt.show()
           # --- Overlap and add ---------------
           xfinal[k:k+len2] = x_old + xi[0:len1]
           x_old = xi[len1:frame_len]
           k = k + len2
   
       # write output
       sf.write(outfile, win_gain * xfinal, samp_rate, 'PCM_16')
   
   
   
   def berouti1(snr):
       beta = 0
       if -5.0 <= snr <= 20:
           beta = 3 - snr * 2 / 20
       elif snr < -5.0:
           beta = 4
       elif snr > 20:
           beta = 1
   
       return beta
   
   
   def berouti(snr):
       beta = 0
       if -5.0 <= snr <= 20:
           beta = 4 - snr * 3 / 20
       elif snr < -5.0:
           beta = 5
       elif snr > 20:
           beta = 1
   
       return beta
   
   
   def next_pow2(n):
       return np.ceil(np.log2(np.abs(n)))
   
   
   def load_audio(filename, trace=0):
       """
       load wav file using audioread.
       This is not available in python x,y.
       """
       data = np.array([])
       with audioread.audio_open(filename) as af:
           trace_n = af.channels
           if trace >= trace_n:
               print('number of traces in file is', trace_n)
               quit()
           nsamp = np.ceil(af.samplerate * af.duration)
           print(f"nsamp =%d" % nsamp)
           data = np.ascontiguousarray(np.zeros(nsamp, 1))
           index = 0
           for buffer in af:
               full_data = np.fromstring(buffer).reshape(-1, af.channels)
               n = full_data.shape[0]
               if index + n > len(data):
                   n = len(data) - index
               if n > 0:
                   data[index:index + n] = full_data[:n, trace]
                   index += n
               else:
                   break
   
       return af.samplerate, data / 2.0 ** 15, 'a.u.'
   
   
   
   
   
   

 

上面代码是根据speech enhancement中的MATLAB代码改写的，仅用作学习用途。

原始的MATLAB代码如下：

function specsub(filename,outfile)

%  Implements the basic power spectral subtraction algorithm [1].
% 
%  Usage:  specsub(noisyFile, outputFile)
%           
%         noisyFile - noisy speech file in .wav format
%         outputFile - enhanced output file in .wav format
%
%   Algorithm uses the first 5 frames to estimate the noise psd, and 
%   then uses a very simple VAD algorithm for updating the noise psd.
%   Alternatively, the VAD in the file ss_rdc.m (this folder) can be used.
%
%  References:
%   [1] Berouti, M., Schwartz, M., and Makhoul, J. (1979). Enhancement of speech 
%       corrupted by acoustic noise. Proc. IEEE Int. Conf. Acoust., Speech, 
%       Signal Processing, 208-211.
%
% Author: Philipos C. Loizou
%
% Copyright (c) 2006 by Philipos C. Loizou
% $Revision: 0.0 $  $Date: 10/09/2006 $
%-------------------------------------------------------------------------

if nargin<2
   fprintf('Usage: specsub noisyfile.wav outFile.wav \n\n');
   return;
end

[x,Srate]=audioread(filename);


% =============== Initialize variables ===============
%

len=floor(20*Srate/1000); % Frame size in samples
if rem(len,2)==1, len=len+1; end;
PERC=50; % window overlap in percent of frame size
len1=floor(len*PERC/100);
len2=len-len1; 


Thres=3; % VAD threshold in dB SNRseg 
alpha=2.0; % power exponent
FLOOR=0.002;
G=0.9;

win=hanning(len, 'periodic'); %tukey(len,PERC);  % define window
winGain=len2/sum(win); % normalization gain for overlap+add with 50% overlap

% Noise magnitude calculations - assuming that the first 5 frames is noise/silence
%
nFFT=2*2^nextpow2(len);
noise_mean=zeros(nFFT,1);
j=1;
for k=1:5
   windowed = win.*x(j:j+len-1);
   noise_mean=noise_mean+abs(fft(win.*x(j:j+len-1),nFFT));
   j=j+len;
end
noise_mu=noise_mean/5;

%--- allocate memory and initialize various variables
   
k=1;
img=sqrt(-1);
x_old=zeros(len1,1);
Nframes=floor(length(x)/len2)-1;
xfinal=zeros(Nframes*len2,1);

%===============================  Start Processing ==================================
%
for n=1:Nframes 
   t = x(k:k+len - 1);
   insign=win.*x(k:k+len-1);     %Windowing  
   spec=fft(insign,nFFT);     %compute fourier transform of a frame
   sig=abs(spec); % compute the magnitude
   
   %save the noisy phase information 
   theta=angle(spec);  
   
   SNRseg=10*log10(norm(sig,2)^2/norm(noise_mu,2)^2);
   
   if alpha==1.0
      beta=berouti1(SNRseg);
   else
     beta=berouti(SNRseg);
  end
   
   
   %&&&&&&&&&
   sub_speech=sig.^alpha - beta*noise_mu.^alpha;
   diffw = sub_speech-FLOOR*noise_mu.^alpha;
   
   % Floor negative components
   z=find(diffw <0);  
   if~isempty(z)
      sub_speech(z)=FLOOR*noise_mu(z).^alpha;
   end
   
   
   % --- implement a simple VAD detector --------------
   %
   if (SNRseg < Thres)   % Update noise spectrum
      noise_temp = G*noise_mu.^alpha+(1-G)*sig.^alpha;   
      noise_mu=noise_temp.^(1/alpha);   % new noise spectrum
   end
   
   
   sub_speech(nFFT/2+2:nFFT)=flipud(sub_speech(2:nFFT/2));  % to ensure conjugate symmetry for real reconstruction
  
   x_phase=(sub_speech.^(1/alpha)).*(cos(theta)+img*(sin(theta)));
  
   
   % take the IFFT 
   xi=real(ifft(x_phase));
       
   % --- Overlap and add ---------------
   xfinal(k:k+len2-1)=x_old+xi(1:len1);
   x_old=xi(1+len1:len);
  
  

 k=k+len2;
end
%========================================================================================


% wavwrite(winGain*xfinal,Srate,16,outfile);
audiowrite(outfile, winGain*xfinal,Srate);

%--------------------------------------------------------------------------

function a=berouti1(SNR)

if SNR>=-5.0 & SNR<=20
   a=3-SNR*2/20;
else
   
  if SNR<-5.0
   a=4;
  end

  if SNR>20
    a=1;
  end
  
end

function a=berouti(SNR)

if SNR>=-5.0 & SNR<=20
   a=4-SNR*3/20; 
else
   
  if SNR<-5.0
   a=5;
  end

  if SNR>20
    a=1;
  end
  
end