add support for WSJ corpus

README.md

@@ -13,7 +13,8 @@ End-to-end automatic speech recognition system implemented in TensorFlow.
 - [x] **Add simple n-gram model for random generation or statistical use** (2017-03-23)
 - [x] **Improve some code for pre-processing and training** (2017-03-23)
 - [x] **Replace TABs with blanks and add nist2wav converter script** (2017-04-20)
-- [x] **Add some data preparation code** (2017-05-1)
+- [x] **Add some data preparation code** (2017-05-01)
+- [x] **Add WSJ corpus standard preprocessing by s5 recipe** (2017-05-05)
 
 ## Recommendation
 If you want to replace feed dict operation with Tensorflow multi-thread and fifoqueue input pipeline, you can refer to my repo [TensorFlow-Input-Pipeline](https://github.com/zzw922cn/TensorFlow-Input-Pipeline) for more example codes. My own practices prove that fifoqueue input pipeline would improve the training speed in some time.

src/feature/core/__init__.py

@@ -0,0 +1,22 @@
+#-*- coding:utf-8 -*-
+#!/usr/bin/python
+''' Automatic Speech Recognition
+author:
+
+      iiiiiiiiiiii            iiiiiiiiiiii         !!!!!!!             !!!!!!    
+      #        ###            #        ###           ###        I#        #:     
+      #      ###              #      I##;             ##;       ##       ##      
+            ###                     ###               !##      ####      #       
+           ###                     ###                 ###    ## ###    #'       
+         !##;                    `##%                   ##;  ##   ###  ##        
+        ###                     ###                     $## `#     ##  #         
+       ###        #            ###        #              ####      ####;         
+     `###        -#           ###        `#               ###       ###          
+     ##############          ##############               `#         #     
+     
+date:2017-4-15
+'''
+
+import sys
+sys.path.append('../')
+sys.dont_write_bytecode = True

src/feature/core/calcmfcc.py

@@ -0,0 +1,183 @@
+#-*- coding:utf-8 -*-
+#!/usr/bin/python
+
+''' calculate mfcc feature vectors.
+author:
+
+      iiiiiiiiiiii            iiiiiiiiiiii         !!!!!!!             !!!!!!    
+      #        ###            #        ###           ###        I#        #:     
+      #      ###              #      I##;             ##;       ##       ##      
+            ###                     ###               !##      ####      #       
+           ###                     ###                 ###    ## ###    #'       
+         !##;                    `##%                   ##;  ##   ###  ##        
+        ###                     ###                     $## `#     ##  #         
+       ###        #            ###        #              ####      ####;         
+     `###        -#           ###        `#               ###       ###          
+     ##############          ##############               `#         #     
+     
+date:2016-11-09
+'''
+
+
+import numpy
+from sigprocess import audio2frame
+from sigprocess import pre_emphasis
+from sigprocess import spectrum_power
+from scipy.fftpack import dct 
+#首先，为了适配版本3.x，需要调整xrange的使用，因为对于版本2.x只能使用range，需要将xrange替换为range
+try:
+    xrange(1)
+except:
+    xrange=range
+
+
+
+def calcMFCC_delta_delta(signal,samplerate=16000,win_length=0.025,win_step=0.01,cep_num=13,filters_num=26,NFFT=512,low_freq=0,high_freq=None,pre_emphasis_coeff=0.97,cep_lifter=22,appendEnergy=True):
+    '''计算13个MFCC+13个一阶微分系数+13个加速系数,一共39个系数
+    '''
+    feat=calcMFCC(signal,samplerate,win_length,win_step,cep_num,filters_num,NFFT,low_freq,high_freq,pre_emphasis_coeff,cep_lifter,appendEnergy)   #首先获取13个一般MFCC系数
+    feat_delta=delta(feat)
+    feat_delta_delta=delta(feat_delta)
+
+    result=numpy.concatenate((feat,feat_delta,feat_delta_delta),axis=1)
+    return result
+
+def delta(feat, N=2):
+    """Compute delta features from a feature vector sequence.
+
+    :param feat: A numpy array of size (NUMFRAMES by number of features) containing features. Each row holds 1 feature vector.
+    :param N: For each frame, calculate delta features based on preceding and following N frames
+    :returns: A numpy array of size (NUMFRAMES by number of features) containing delta features. Each row holds 1 delta feature vector.
+    """
+    NUMFRAMES = len(feat)
+    feat = numpy.concatenate(([feat[0] for i in range(N)], feat, [feat[-1] for i in range(N)]))
+    denom = sum([2*i*i for i in range(1,N+1)])
+    dfeat = []
+    for j in range(NUMFRAMES):
+        dfeat.append(numpy.sum([n*feat[N+j+n] for n in range(-1*N,N+1)], axis=0)/denom)
+    return dfeat
+
+def calcMFCC(signal,samplerate=16000,win_length=0.025,win_step=0.01,cep_num=13,filters_num=26,NFFT=512,low_freq=0,high_freq=None,pre_emphasis_coeff=0.97,cep_lifter=22,appendEnergy=True):
+    '''计算13个MFCC系数
+    参数含义：
+    signal:原始音频信号，一般为.wav格式文件
+    samplerate:抽样频率，这里默认为16KHz
+    win_length:窗长度，默认即一帧为25ms
+    win_step:窗间隔，默认情况下即相邻帧开始时刻之间相隔10ms
+    cep_num:倒谱系数的个数，默认为13
+    filters_num:滤波器的个数，默认为26
+    NFFT:傅立叶变换大小，默认为512
+    low_freq:最低频率，默认为0
+    high_freq:最高频率
+    pre_emphasis_coeff:预加重系数，默认为0.97
+    cep_lifter:倒谱的升个数
+    appendEnergy:是否加上能量，默认加
+    '''
+
+    feat,energy=fbank(signal,samplerate,win_length,win_step,filters_num,NFFT,low_freq,high_freq,pre_emphasis_coeff)
+    feat=numpy.log(feat)
+    feat=dct(feat,type=2,axis=1,norm='ortho')[:,:cep_num]  #进行离散余弦变换,只取前13个系数
+    feat=lifter(feat,cep_lifter)
+    if appendEnergy:
+	feat[:,0]=numpy.log(energy)  #只取2-13个系数，第一个用能量的对数来代替
+    return feat
+
+def fbank(signal,samplerate=16000,win_length=0.025,win_step=0.01,filters_num=26,NFFT=512,low_freq=0,high_freq=None,pre_emphasis_coeff=0.97):
+    '''计算音频信号的MFCC
+    参数说明：
+    samplerate:采样频率
+    win_length:窗长度
+    win_step:窗间隔
+    filters_num:梅尔滤波器个数
+    NFFT:FFT大小
+    low_freq:最低频率
+    high_freq:最高频率
+    pre_emphasis_coeff:预加重系数
+    '''
+
+    high_freq=high_freq or samplerate/2  #计算音频样本的最大频率
+    signal=pre_emphasis(signal,pre_emphasis_coeff)  #对原始信号进行预加重处理
+    frames=audio2frame(signal,win_length*samplerate,win_step*samplerate) #得到帧数组
+    spec_power=spectrum_power(frames,NFFT)  #得到每一帧FFT以后的功率谱
+    energy=numpy.sum(spec_power,1)  #对每一帧的功率谱进行求和，得到能量
+    energy=numpy.where(energy==0,numpy.finfo(float).eps,energy)  #对能量为0的地方调整为eps，这样便于进行对数处理
+    fb=get_filter_banks(filters_num,NFFT,samplerate,low_freq,high_freq)  #获得每一个滤波器的频率宽度
+    feat=numpy.dot(spec_power,fb.T)  #对滤波器和功率谱进行点乘
+    feat=numpy.where(feat==0,numpy.finfo(float).eps,feat)  #同样不能出现0
+    return feat,energy
+
+def log_fbank(signal,samplerate=16000,win_length=0.025,win_step=0.01,filters_num=26,NFFT=512,low_freq=0,high_freq=None,pre_emphasis_coeff=0.97):
+    '''计算对数值
+    参数含义：同上
+    '''
+    feat,energy=fbank(signal,samplerate,win_length,win_step,filters_num,NFFT,low_freq,high_freq,pre_emphasis_coeff)
+    return numpy.log(feat)
+
+def ssc(signal,samplerate=16000,win_length=0.025,win_step=0.01,filters_num=26,NFFT=512,low_freq=0,high_freq=None,pre_emphasis_coeff=0.97):
+    '''
+    待补充
+    ''' 
+    high_freq=high_freq or samplerate/2
+    signal=sigprocess.pre_emphasis(signal,pre_emphasis_coeff)
+    frames=sigprocess.audio2frame(signal,win_length*samplerate,win_step*samplerate)
+    spec_power=sigprocess.spectrum_power(frames,NFFT) 
+    spec_power=numpy.where(spec_power==0,numpy.finfo(float).eps,spec_power) #能量谱
+    fb=get_filter_banks(filters_num,NFFT,samplerate,low_freq,high_freq) 
+    feat=numpy.dot(spec_power,fb.T)  #计算能量
+    R=numpy.tile(numpy.linspace(1,samplerate/2,numpy.size(spec_power,1)),(numpy.size(spec_power,0),1))
+    return numpy.dot(spec_power*R,fb.T)/feat
+
+def hz2mel(hz):
+    '''把频率hz转化为梅尔频率
+    参数说明：
+    hz:频率
+    '''
+    return 2595*numpy.log10(1+hz/700.0)
+
+def mel2hz(mel):
+    '''把梅尔频率转化为hz
+    参数说明：
+    mel:梅尔频率
+    '''
+    return 700*(10**(mel/2595.0)-1)
+
+def get_filter_banks(filters_num=20,NFFT=512,samplerate=16000,low_freq=0,high_freq=None):
+    '''计算梅尔三角间距滤波器，该滤波器在第一个频率和第三个频率处为0，在第二个频率处为1
+    参数说明：
+    filers_num:滤波器个数
+    NFFT:FFT大小
+    samplerate:采样频率
+    low_freq:最低频率
+    high_freq:最高频率
+    '''
+    #首先，将频率hz转化为梅尔频率，因为人耳分辨声音的大小与频率并非线性正比，所以化为梅尔频率再线性分隔
+    low_mel=hz2mel(low_freq)
+    high_mel=hz2mel(high_freq)
+    #需要在low_mel和high_mel之间等间距插入filters_num个点，一共filters_num+2个点
+    mel_points=numpy.linspace(low_mel,high_mel,filters_num+2)
+    #再将梅尔频率转化为hz频率，并且找到对应的hz位置
+    hz_points=mel2hz(mel_points)
+    #我们现在需要知道这些hz_points对应到fft中的位置
+    bin=numpy.floor((NFFT+1)*hz_points/samplerate)
+    #接下来建立滤波器的表达式了，每个滤波器在第一个点处和第三个点处均为0，中间为三角形形状
+    fbank=numpy.zeros([filters_num,NFFT/2+1])
+    for j in xrange(0,filters_num):
+	for i in xrange(int(bin[j]),int(bin[j+1])):
+	    fbank[j,i]=(i-bin[j])/(bin[j+1]-bin[j])
+	for i in xrange(int(bin[j+1]),int(bin[j+2])):
+	    fbank[j,i]=(bin[j+2]-i)/(bin[j+2]-bin[j+1])
+    return fbank
+
+def lifter(cepstra,L=22):
+    '''升倒谱函数
+    参数说明：
+    cepstra:MFCC系数
+    L：升系数，默认为22
+    '''
+    if L>0:
+	nframes,ncoeff=numpy.shape(cepstra)
+	n=numpy.arange(ncoeff)
+	lift=1+(L/2)*numpy.sin(numpy.pi*n/L)
+	return lift*cepstra
+    else:
+	return cepstra

src/feature/core/fileUtils.py

@@ -0,0 +1,21 @@
+#-*- coding:utf-8 -*-
+#!/usr/bin/python
+''' Automatic Speech Recognition
+
+author(s):
+zzw922cn
+     
+date:2017-5-5
+'''
+
+import sys
+sys.path.append('../')
+sys.dont_write_bytecode = True
+
+import os
+
+def check_path_exists(path):
+  """ check a path exists or not
+  """
+  if not os.path.exists(path):
+    os.makedirs(path)

src/feature/core/nist2wav.py

@@ -0,0 +1,20 @@
+#!/usr/bin/python
+
+import sys
+sys.path.append('../')
+sys.dont_write_bytecode = True
+
+import subprocess
+import os
+
+def nist2wav(src_dir):
+  count = 0
+  for subdir, dirs, files in os.walk(src_dir):
+    for f in files:
+      fullFilename = os.path.join(subdir, f)
+      if f.endswith('.wv1') or f.endswith('.wv2'):
+        count += 1
+        os.system("./sph2pipe_v2.5/sph2pipe "+fullFilename+" -f rif " +fullFilename+".wav")
+        print fullFilename
+
+nist2wav('/home/pony/wsj/')

src/feature/core/nist2wav.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+# author:Rongyi Lan
+# email:brianlanbo@gmail.com
+
+# This file is designed for converting NIST format audio
+# to WAV format audio, to run this script, you should install
+# libsndfile software first.
+
+
+target_dir=$1
+
+fnames=(`find $target_dir -name "*.wv1"`)
+
+for fname in "${fnames[@]}"
+do
+  mv "$fname" "${fname%.wav}.nist"
+  sndfile-convert "${fname%.wav}.nist" "$fname"
+  if [ $? = 0 ]; then
+    echo renamed $fname to nist and converted back to wav using sndfile-convert
+  else
+    mv "${fname%.wav}.nist" "$fname"
+  fi
+done