最小语音识别模型
- 输入
- 音频特征向量，共计26维的mfcc特征向量
- 为了使神经网络能够统一格式，将[batch_size, time_step, 26]中的time_step定为一个较大的值，如果实际时长不足，采用补零的方法补齐
- 输出
- 输出为一个[batch_size, words_length]的张量，方便训练将word_length也取一个较大的常量，不足的补零
- 网络用keras进行训练，利用backend设置ctc。
模型训练
import os
import numpy as np
import scipy.io.wavfile as wav
from collections import Counter
from python_speech_features import mfcc
from keras.models import Model
from keras.layers import Dense, Dropout, Input
from keras.layers import Lambda, Activation
from keras.layers.merge import add, concatenate
from keras import backend as K
from keras.optimizers import SGD, Adadelta
from keras.layers.recurrent import GRU
from keras.preprocessing.sequence import pad_sequences
# 获取音频文件列表及音频id
def genwavlist(wavpath):
wavfiles = {}
fileids = []
for (dirpath, dirnames, filenames) in os.walk(wavpath):
for filename in filenames:
if filename.endswith('.wav'):
filepath = os.sep.join([dirpath, filename])
fileid = filename.strip('.wav')
wavfiles[fileid] = filepath
fileids.append(fileid)
return wavfiles,fileids
# 计算mfcc，并将特征补零为[500,26]的shape
def compute_mfcc(file):
fs, audio = wav.read(file)
mfcc_feat = mfcc(audio, samplerate=fs, numcep=26)
mfcc_feat = mfcc_feat[::3]
mfcc_feat = np.transpose(mfcc_feat)
mfcc_feat = pad_sequences(mfcc_feat, maxlen=500, dtype='float', padding='post', truncating='post').T
return mfcc_feat
# 生成拼音映射到符号的map
def gendict(textfile_path):
dicts = []
textfile = open(textfile_path,'r+')
for content in textfile.readlines():
content = content.strip('\n')
content = content.split(' ',1)[1]
content = content.split(' ')
dicts += (word for word in content)
counter = Counter(dicts)
words = sorted(counter)
wordsize = len(words)
word2num = dict(zip(words, range(wordsize)))
return word2num,len(word2num)
# 利用字典，将text映射为number
def text2num(textfile_path):
lexcion,wordnum = gendict(textfile_path)
word2num = lambda word:lexcion.get(word, 0)
textfile = open(textfile_path, 'r+')
content_dict = {}
for content in textfile.readlines():
content = content.strip('\n')
cont_id = content.split(' ',1)[0]
content = content.split(' ',1)[1]
content = content.split(' ')
content = list(map(word2num,content))
add_num = list(np.zeros(50-len(content)))
content = content + add_num
content_dict[cont_id] = content
return content_dict,lexcion
# 将MFCC和number整理为能够被模型所训练的格式
def get_batch(x, y, train=False, max_pred_len=50, input_length=500):
X = np.expand_dims(x, axis=3)
X = x # for model2
# labels = np.ones((y.shape[0], max_pred_len)) * -1 # 3 # , dtype=np.uint8
labels = y
input_length = np.ones([x.shape[0], 1]) * ( input_length - 2 )
# label_length = np.ones([y.shape[0], 1])
label_length = np.sum(labels > 0, axis=1)
label_length = np.expand_dims(label_length,1)
inputs = {'the_input': X,
'the_labels': labels,
'input_length': input_length,
'label_length': label_length,
}
outputs = {'ctc': np.zeros([x.shape[0]])} # dummy data for dummy loss function
return (inputs, outputs)
# 训练数据的生成器
def data_generate(wavpath = 'D:\\workspace\\github\\data', textfile = 'D:\\workspace\\github\\data\\test.txt', bath_size=1):
# 生成音频列表
wavdict,fileids = genwavlist(wavpath)
# 将text转化为number
content_dict,lexcion = text2num(textfile)
# 随意写的循环，测试编写的
genloop = len(fileids)//bath_size
for i in range(genloop):
print("the ",i,"'s loop")
feats = []
labels = []
for x in range(bath_size):
num = i * bath_size + x
fileid = fileids[num]
# 生成特征
mfcc_feat = compute_mfcc(wavdict[fileid])
# 一个batch的特征被压进来了
feats.append(mfcc_feat)
# 一个batch的label被压进来了
labels.append(content_dict[fileid])
feats = np.array(feats)
labels = np.array(labels)
# 整理数据格式
inputs, outputs = get_batch(feats, labels)
# 利用yield代替return，是生成器的特殊用法
yield inputs, outputs
# 利用backend调用ctc
def ctc_lambda(args):
labels, y_pred, input_length, label_length = args
y_pred = y_pred[:, :, :]
return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
def creatModel():
input_data = Input(name='the_input', shape=(500, 26))
layer_h1 = Dense(512, activation="relu", use_bias=True, kernel_initializer='he_normal')(input_data)
layer_h1 = Dropout(0.3)(layer_h1)
layer_h2 = Dense(512, activation="relu", use_bias=True, kernel_initializer='he_normal')(layer_h1)
layer_h3_1 = GRU(512, return_sequences=True, kernel_initializer='he_normal', dropout=0.3)(layer_h2) # GRU
layer_h3_2 = GRU(512, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', dropout=0.3)(layer_h2) # GRU
layer_h3 = add([layer_h3_1, layer_h3_2])
layer_h4 = Dense(512, activation="relu", use_bias=True, kernel_initializer='he_normal')(layer_h3)
layer_h4 = Dropout(0.3)(layer_h4)
layer_h5 = Dense(1200, activation="relu", use_bias=True, kernel_initializer='he_normal')(layer_h4)
output = Activation('softmax', name='Activation0')(layer_h5)
model_data = Model(inputs=input_data, outputs=output)
#ctc
labels = Input(name='the_labels', shape=[50], dtype='float32')
input_length = Input(name='input_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
loss_out = Lambda(ctc_lambda, output_shape=(1,), name='ctc')([labels, output, input_length, label_length])
model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)
model.summary()
ada_d = Adadelta(lr=0.01, rho=0.95, epsilon=1e-06)
model.compile(loss={'ctc': lambda y_true, output: output}, optimizer=ada_d)
#test_func = K.function([input_data], [output])
print("model compiled successful!")
return model, model_data
model, model_data = creatModel()
# 定义数据生成器
yielddatas = data_generate()
# 使用fit_generator进行训练
model.fit_generator(yielddatas,1)
model.save_weights('model.mdl')
model_data.save_weights('model_data.mdl')
#text2num('E:\\Data\\thchs30\\cv.syllable.txt')后续
等训练好了试试识别怎么样，目前还没训练。
用于记录学习，希望大家提出宝贵意见