TensorFlow练习15: 中文语音识别

语音识别的应用领域非常广泛，洋文名Speech Recognition。它所要解决的问题是让计算机能够“听懂”人类的语音，将语音中包含的文字信息“提取”出来。

语音识别是前文《聊天机器人》必不可少的一个组件，本帖就使用TensorFlow做一个中文语音识别。

使用的数据集

THCHS30是Dong Wang, Xuewei Zhang, Zhiyong Zhang这几位大神发布的开放语音数据集，可用于开发中文语音识别系统。

为了感谢这几位大神，我是跪在电脑前写的本帖代码。

下载中文语音数据集（5G+）：

1 2 3 4 5 6 7

$ wget http : / / data .cslt .org / thchs30 / zip / wav .tgz $ wget http : / / data .cslt .org / thchs30 / zip / doc .tgz $ wget http : / / data .cslt .org / thchs30 / zip / lm .tgz # 解压 $ tar xvf wav .tgz $ tar xvf doc .tgz $ tar xvf lm .tgz

在开始之前，先好好检视一下数据集。

训练

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266

import tensorflow as tf    # 0.12 import numpy as np import os from collections import Counter import librosa    # https://github.com/librosa/librosa   # 训练样本路径 wav_path = 'data/wav/train' label_file = 'data/doc/trans/train.word.txt'   # 获得训练用的wav文件路径列表 def get_wav_files ( wav_path = wav_path ) : wav_files = [ ] for ( dirpath , dirnames , filenames ) in os . walk ( wav_path ) : for filename in filenames : if filename . endswith ( '.wav' ) or filename . endswith ( '.WAV' ) : filename_path = os . sep . join ( [ dirpath , filename ] ) if os . stat ( filename_path ) . st_size < 240000 :    # 剔除掉一些小文件 continue wav_files . append ( filename_path ) return wav_files   wav_files = get_wav_files ( )   # 读取wav文件对应的label def get_wav_lable ( wav_files = wav_files , label_file = label_file ) : labels_dict = { } with open ( label_file , 'r' ) as f : for label in f : label = label . strip ( '\n' ) label_id = label . split ( ' ' , 1 ) [ 0 ] label_text = label . split ( ' ' , 1 ) [ 1 ] labels_dict [ label_id ] = label_text   labels = [ ] new_wav_files = [ ] for wav_file in wav_files : wav_id = os.path . basename ( wav_file ) . split ( '.' ) [ 0 ] if wav_id in labels_dict : labels . append ( labels_dict [ wav_id ] ) new_wav_files . append ( wav_file )   return new_wav_files , labels   wav_files , labels = get_wav_lable ( ) print ( "样本数:" , len ( wav_files ) )    # 8911 #print(wav_files[0], labels[0]) # wav/train/A11/A11_0.WAV -> 绿 是 阳春 烟 景 大块 文章 的 底色 四月 的 林 峦 更是 绿 得 鲜活 秀媚 诗意 盎然   # 词汇表(参看练习1和7) all_words = [ ] for label in labels : all_words += [ word for word in label ] counter = Counter ( all_words ) count_pairs = sorted ( counter . items ( ) , key = lambda x : - x [ 1 ] )   words , _ = zip ( * count_pairs ) words_size = len ( words ) print ( '词汇表大小:' , words_size )   word_num_map = dict ( zip ( words , range ( len ( words ) ) ) ) to_num = lambda word : word_num_map . get ( word , len ( words ) ) labels_vector = [ list ( map ( to_num , label ) ) for label in labels ] #print(wavs_file[0], labels_vector[0]) #wav/train/A11/A11_0.WAV -> [479, 0, 7, 0, 138, 268, 0, 222, 0, 714, 0, 23, 261, 0, 28, 1191, 0, 1, 0, 442, 199, 0, 72, 38, 0, 1, 0, 463, 0, 1184, 0, 269, 7, 0, 479, 0, 70, 0, 816, 254, 0, 675, 1707, 0, 1255, 136, 0, 2020, 91] #print(words[479]) #绿 label_max_len = np . max ( [ len ( label ) for label in labels_vector ] ) print ( '最长句子的字数:' , label_max_len )   wav_max_len = 0    # 673 for wav in wav_files :      wav , sr = librosa . load ( wav , mono = True )      mfcc = np . transpose ( librosa . feature . mfcc ( wav , sr ) , [ 1 , 0 ] )      if len ( mfcc ) > wav_max_len :          wav_max_len = len ( mfcc ) print ( "最长的语音:" , wav_max_len )   batch_size = 16 n_batch = len ( wav_files ) / / batch_size   # 获得一个batch pointer = 0 def get_next_batches ( batch_size ) :      global pointer      batches_wavs = [ ]      batches_labels = [ ]      for i in range ( batch_size ) :          wav , sr = librosa . load ( wav_files [ pointer ] , mono = True )          mfcc = np . transpose ( librosa . feature . mfcc ( wav , sr ) , [ 1 , 0 ] )          batches_wavs . append ( mfcc . tolist ( ) )          batches_labels . append ( labels_vector [ pointer ] )          pointer += 1        # 补零对齐      for mfcc in batches_wavs :          while len ( mfcc ) < wav_max_len :              mfcc . append ( [ 0 ] * 20 )      for label in batches_labels :          while len ( label ) < label_max_len :              label . append ( 0 )      return batches_wavs , batches _labels   X = tf . placeholder ( dtype = tf . float32 , shape = [ batch_size , None , 20 ] ) sequence_len = tf . reduce_sum ( tf . cast ( tf . not_equal ( tf . reduce_sum ( X , reduction_indices = 2 ) , 0. ) , tf . int32 ) , reduction_indices = 1 ) Y = tf . placeholder ( dtype = tf . int32 , shape = [ batch_size , None ] )   # conv1d_layer conv1d_index = 0 def conv1d_layer ( input_tensor , size , dim , activation , scale , bias ) :      global conv1d_index      with tf . variable_scope ( 'conv1d_' + str ( conv1d_index ) ) :          W = tf . get_variable ( 'W' , ( size , input_tensor . get_shape ( ) . as_list ( ) [ - 1 ] , dim ) , dtype = tf . float32 , initializer = tf . random_uniform_initializer ( minval = - scale , maxval = scale ) )          if bias :              b = tf . get_variable ( 'b' , [ dim ] , dtype = tf . float32 , initializer = tf . constant_initializer ( 0 ) )          out = tf . nn . conv1d ( input_tensor , W , stride = 1 , padding = 'SAME' ) + ( b if bias else 0 )          if not bias :              beta = tf . get_variable ( 'beta' , dim , dtype = tf . float32 , initializer = tf . constant_initializer ( 0 ) )              gamma = tf . get_variable ( 'gamma' , dim , dtype = tf . float32 , initializer = tf . constant_initializer ( 1 ) )              mean_running = tf . get_variable ( 'mean' , dim , dtype = tf . float32 , initializer = tf . constant_initializer ( 0 ) )              variance_running = tf . get_variable ( 'variance' , dim , dtype = tf . float32 , initializer = tf . constant_initializer ( 1 ) )              mean , variance = tf . nn . moments ( out , axes = range ( len ( out . get_shape ( ) ) - 1 ) )              def update_running_stat ( ) :                  decay = 0.99                  update_op = [ mean_running . assign ( mean_running * decay + mean * ( 1 - decay ) ) , variance_running . assign ( variance_running * decay + variance * ( 1 - decay ) ) ]                  with tf . control_dependencies ( update_op ) :                      return tf . identity ( mean ) , tf . identity ( variance )                  m , v = tf . cond ( tf . Variable ( False , trainable = False , collections = [ tf . GraphKeys . LOCAL_VARIABLES ] ) , update_running_stat , lambda : ( mean_running , variance_running ) )                  out = tf . nn . batch_normalization ( out , m , v , beta , gamma , 1e - 8 )          if activation == 'tanh' :              out = tf . nn . tanh ( out )          if activation == 'sigmoid' :              out = tf . nn . sigmoid ( out )            conv1d_index += 1          return out # aconv1d_layer aconv1d_index = 0 def aconv1d_layer ( input_tensor , size , rate , activation , scale , bias ) :      global aconv1d_index      with tf . variable_scope ( 'aconv1d_' + str ( aconv1d_index ) ) :          shape = input_tensor . get_shape ( ) . as_list ( )          W = tf . get_variable ( 'W' , ( 1 , size , shape [ - 1 ] , shape [ - 1 ] ) , dtype = tf . float32 , initializer = tf . random_uniform_initializer ( minval = - scale , maxval = scale ) )          if bias :              b = tf . get_variable ( 'b' , [ shape [ - 1 ] ] , dtype = tf . float32 , initializer = tf . constant_initializer ( 0 ) )          out = tf . nn . atrous_conv2d ( tf . expand_dims ( input_tensor , dim = 1 ) , W , rate = rate , padding = 'SAME' )          out = tf . squeeze ( out , [ 1 ] )          if not bias :              beta = tf . get_variable ( 'beta' , shape [ - 1 ] , dtype = tf . float32 , initializer = tf . constant_initializer ( 0 ) )              gamma = tf . get_variable ( 'gamma' , shape [ - 1 ] , dtype = tf . float32 , initializer = tf . constant_initializer ( 1 ) )              mean_running = tf . get_variable ( 'mean' , shape [ - 1 ] , dtype = tf . float32 , initializer = tf . constant_initializer ( 0 ) )              variance_running = tf . get_variable ( 'variance' , shape [ - 1 ] , dtype = tf . float32 , initializer = tf . constant_initializer ( 1 ) )              mean , variance = tf . nn . moments ( out , axes = range ( len ( out . get_shape ( ) ) - 1 ) )              def update_running_stat ( ) :                  decay = 0.99                  update_op = [ mean_running . assign ( mean_running * decay + mean * ( 1 - decay ) ) , variance_running . assign ( variance_running * decay + variance * ( 1 - decay ) ) ]                  with tf . control_dependencies ( update_op ) :                      return tf . identity ( mean ) , tf . identity ( variance )                  m , v = tf . cond ( tf . Variable ( False , trainable = False , collections = [ tf . GraphKeys . LOCAL_VARIABLES ] ) , update_running_stat , lambda : ( mean_running , variance_running ) )                  out = tf . nn . batch_normalization ( out , m , v , beta , gamma , 1e - 8 )          if activation == 'tanh' :              out = tf . nn . tanh ( out )          if activation == 'sigmoid' :              out = tf . nn . sigmoid ( out )            aconv1d_index += 1          return out # 定义神经网络 def speech_to_text_network ( n_dim = 128 , n_blocks = 3 ) :      out = conv1d_layer ( input_tensor = X , size = 1 , dim = n_dim , activation = 'tanh' , scale = 0.14 , bias = False )      # skip connections      def residual_block ( input_sensor , size , rate ) :              conv_filter = aconv1d_layer ( input_sensor , size = size , rate = rate , activation = 'tanh' , scale = 0.03 , bias = False )              conv_gate = aconv1d_layer ( input_sensor , size = size , rate = rate ,    activation = 'sigmoid' , scale = 0.03 , bias = False )              out = conv_filter * conv_gate              out = conv1d_layer ( out , size = 1 , dim = n_dim , activation = 'tanh' , scale = 0.08 , bias = False )              return out + input_sensor , out      skip = 0      for _ in range ( n_blocks ) :          for r in [ 1 , 2 , 4 , 8 , 16 ] :              out , s = residual_block ( out , size = 7 , rate = r )              skip += s        logit = conv1d_layer ( skip , size = 1 , dim = skip . get_shape ( ) . as_list ( ) [ - 1 ] , activation = 'tanh' , scale = 0.08 , bias = False )      logit = conv1d_layer ( logit , size = 1 , dim = words_size , activation = None , scale = 0.04 , bias = True )        return logit   class MaxPropOptimizer ( tf . train . Optimizer ) :      def __init__ ( self , learning_rate = 0.001 , beta2 = 0.999 , use_locking = False , name = "MaxProp" ) :          super ( MaxPropOptimizer , self ) . __init__ ( use_locking , name )          self . _lr = learning_rate          self . _beta2 = beta2          self . _lr_t = None          self . _beta2_t = None      def _prepare ( self ) :          self . _lr_t = tf . convert_to_tensor ( self . _lr , name = "learning_rate" )          self . _beta2_t = tf . convert_to_tensor ( self . _beta2 , name = "beta2" )      def _create_slots ( self , var_list ) :          for v in var_list :              self . _zeros_slot ( v , "m" , self . _name )      def _apply_dense ( self , grad , var ) :          lr_t = tf . cast ( self . _lr_t , var . dtype . base_dtype )          beta2_t = tf . cast ( self . _beta2_t , var . dtype . base_dtype )          if var . dtype . base_dtype == tf . float16 :              eps = 1e - 7          else :              eps = 1e - 8          m = self . get_slot ( var , "m" )          m_t = m . assign ( tf . maximum ( beta2_t * m + eps , tf . abs ( grad ) ) )          g_t = grad / m_t          var_update = tf . assign_sub ( var , lr_t * g_t )          return tf . group ( * [ var_update , m_t ] )      def _apply_sparse ( self , grad , var ) :          return self . _apply_dense ( grad , var )   def train_speech_to_text_network ( ) :      logit = speech_to_text_network ( )        # CTC loss      indices = tf . where ( tf . not_equal ( tf . cast ( Y , tf . float32 ) , 0. ) )      target = tf . SparseTensor ( indices = indices , values = tf . gather_nd ( Y , indices ) - 1 , shape = tf . cast ( tf . shape ( Y ) , tf . int64 ) )      loss = tf . nn . ctc_loss ( logit , target , sequence_len , time_major = False )      # optimizer      lr = tf . Variable ( 0.001 , dtype = tf . float32 , trainable = False )      optimizer = MaxPropOptimizer ( learning_rate = lr , beta2 = 0.99 )      var_list = [ t for t in tf . trainable_variables ( ) ]      gradient = optimizer . compute_gradients ( loss , var_list = var_list )      optimizer_op = optimizer . apply_gradients ( gradient )        with tf . Session ( ) as sess :          sess . run ( tf . global_variables_initializer ( ) )            saver = tf . train . Saver ( tf . global_variables ( ) )            for epoch in range ( 16 ) :              sess . run ( tf . assign ( lr , 0.001 * ( 0.97 * * epoch ) ) )                global pointer              pointer = 0              for batch in range ( n_batch ) :                  batches_wavs , batches_labels = get_next_batches ( batch_size )                  train_loss , _ = sess . run ( [ loss , optimizer_op ] , feed_dict = { X : batches_wavs , Y : batches_labels } )                  print ( epoch , batch , train_loss )              if epoch % 5 == 0 :                  saver . save ( sess , 'speech.module' , global_step = epoch )   # 训练 train_speech_to_text_network ( )   # 语音识别 # 把batch_size改为1 def speech_to_text ( wav_file ) :      wav , sr = librosa . load ( wav_file , mono = True )      mfcc = np . transpose ( np . expand_dims ( librosa . feature . mfcc ( wav , sr ) , axis = 0 ) , [ 0 , 2 , 1 ] )        logit = speech_to_text_network ( )        saver = tf . train . Saver ( )      with tf . Session ( ) as sess :          saver . restore ( sess , tf . train . latest_checkpoint ( '.' ) )            decoded = tf . transpose ( logit , perm = [ 1 , 0 , 2 ] )          decoded , _ = tf . nn . ctc_beam_search_decoder ( decoded , sequence_len , merge_repeated = False )          predict = tf . sparse_to_dense ( decoded [ 0 ] . indices , decoded [ 0 ] . shape , decoded [ 0 ] . values ) + 1          output = sess . run ( decoded , feed_dict = { X : mfcc } )                   #print(output)

相关资源：

• TensorFlow练习8: 基于RNN生成音乐
• Machine Learning is Fun Part 6: How to do Speech Recognition with Deep Learning
• 深度学习大牛Andrew Ng：Speech Recognition and Beyond
• https://github.com/kaldi-asr/kaldi
• http://cmusphinx.sourceforge.net
• https://pypi.python.org/pypi/SpeechRecognition