基本完成除添加模型之外的其他部分,不过尚未测试
This commit is contained in:
parent
c621c0aeb9
commit
4ca877a046
44
README.md
44
README.md
|
@ -1,43 +1 @@
|
|||
# ASRT_SpeechRecognition
|
||||
基于深度学习的语音识别系统
|
||||
|
||||
## Introduction
|
||||
简介
|
||||
|
||||
可以尝试使用Keras进行制作
|
||||
|
||||
本项目将使用TensorFlow基于递归神经网络和卷积神经网络进行制作。
|
||||
|
||||
This project will use TensorFlow based on RNN and CNN to implement.
|
||||
|
||||
本项目尚未完成,想要Fork的同学请手慢。
|
||||
|
||||
## Model
|
||||
模型
|
||||
|
||||
### Speech Model
|
||||
语音模型
|
||||
|
||||
LSTM + CNN
|
||||
|
||||
### Language Model
|
||||
语言模型
|
||||
|
||||
基于概率图的马尔可夫模型
|
||||
|
||||
## Python Import
|
||||
Python的依赖库
|
||||
|
||||
* python_speech_features
|
||||
* TensorFlow
|
||||
* Keras
|
||||
* Numpy
|
||||
* wave
|
||||
* matplotlib
|
||||
* math
|
||||
* Scipy
|
||||
|
||||
## Log
|
||||
日志
|
||||
|
||||
链接:[进展日志](https://github.com/nl8590687/ASRT_SpeechRecognition/blob/master/log.md)
|
||||
# ASRT_SpeechRecognition
基于深度学习的语音识别系统
## Introduction 简介
本项目使用Keras、TensorFlow基于长短时记忆神经网络和卷积神经网络以及CTC进行制作。
This project uses keras, TensorFlow based on LSTM, CNN and CTC to implement.
本项目尚未完成,想要Fork的同学请手慢。
## Model 模型
### Speech Mode l语音模型
CNN + LSTM + CTC
### Language Model 语言模型
基于概率图的马尔可夫模型
## Python Import
Python的依赖库
* python_speech_features
* TensorFlow
* Keras
* Numpy
* wave
* matplotlib
* math
* Scipy
## Log
日志
链接:[进展日志](https://github.com/nl8590687/ASRT_SpeechRecognition/blob/master/log.md)
|
19
log.md
19
log.md
|
@ -1,18 +1 @@
|
|||
# ASRT_SpeechRecognition
|
||||
基于深度学习的语音识别系统
|
||||
|
||||
## Introduction
|
||||
|
||||
这里是更新记录日志文件
|
||||
|
||||
如果有什么问题,团队内部需要在这里直接写出来
|
||||
|
||||
## Log
|
||||
### 2017-08-31
|
||||
数据处理部分的代码基本完成,现在准备撸模型
|
||||
### 2017-08-29
|
||||
准备使用现有的包[python_speech_features](https://github.com/jameslyons/python_speech_features)来实现特征的提取,以及求一阶二阶差分。
|
||||
### 2017-08-28
|
||||
开始准备制作语音信号处理方面的功能
|
||||
### 2017-08-22
|
||||
准备使用Keras基于LSTM/CNN尝试实现
|
||||
# ASRT_SpeechRecognition
基于深度学习的语音识别系统
## Introduction
这里是更新记录日志文件
如果有什么问题,团队内部需要在这里直接写出来
## Log
### 2017-09-08
基本完成除了添加模型之外的其他部分代码
### 2017-08-31
数据处理部分的代码基本完成,现在准备撸模型
### 2017-08-29
准备使用现有的包[python_speech_features](https://github.com/jameslyons/python_speech_features)来实现特征的提取,以及求一阶二阶差分。
### 2017-08-28
开始准备制作语音信号处理方面的功能
### 2017-08-22
准备使用Keras基于LSTM/CNN尝试实现
|
57
main.py
57
main.py
|
@ -15,11 +15,11 @@ from keras import backend as K
|
|||
from readdata import DataSpeech
|
||||
|
||||
class ModelSpeech(): # 语音模型类
|
||||
def __init__(self,MS_EMBED_SIZE = 64,BATCH_SIZE = 32):
|
||||
def __init__(self,MS_OUTPUT_SIZE = 1283,BATCH_SIZE = 32):
|
||||
'''
|
||||
初始化
|
||||
'''
|
||||
self.MS_EMBED_SIZE = MS_EMBED_SIZE # LSTM 的大小
|
||||
self.MS_OUTPUT_SIZE = MS_OUTPUT_SIZE # 神经网络最终输出的每一个字符向量维度的大小
|
||||
self.BATCH_SIZE = BATCH_SIZE # 一次训练的batch
|
||||
self._model = self.CreateModel()
|
||||
|
||||
|
@ -29,10 +29,13 @@ class ModelSpeech(): # 语音模型类
|
|||
输入层:39维的特征值序列,一条语音数据的最大长度设为1500(大约15s)
|
||||
隐藏层一:1024个神经元的卷积层
|
||||
隐藏层二:池化层,池化窗口大小为2
|
||||
隐藏层三:Dropout层,需要断开的神经元的比例为0.3,防止过拟合
|
||||
隐藏层三:Dropout层,需要断开的神经元的比例为0.2,防止过拟合
|
||||
隐藏层四:循环层、LSTM层
|
||||
隐藏层五:Dropout层,需要断开的神经元的比例为0.3,防止过拟合
|
||||
输出层:全连接层,神经元数量为1279,使用softmax作为激活函数,使用CTC的loss作为损失函数
|
||||
隐藏层五:Dropout层,需要断开的神经元的比例为0.2,防止过拟合
|
||||
隐藏层六:全连接层,神经元数量为self.MS_OUTPUT_SIZE,使用softmax作为激活函数,
|
||||
输出层:lambda层,即CTC层,使用CTC的loss作为损失函数,实现多输出
|
||||
|
||||
当前未完成,针对多输出的CTC层尚未添加
|
||||
'''
|
||||
# 每一帧使用13维mfcc特征及其13维一阶差分和13维二阶差分表示,最大信号序列长度为1500
|
||||
layer_input = Input((1500,39))
|
||||
|
@ -42,23 +45,12 @@ class ModelSpeech(): # 语音模型类
|
|||
layer_h3 = Dropout(0.2)(layer_h2) # 随机中断部分神经网络连接,防止过拟合
|
||||
layer_h4 = LSTM(256, activation='relu', use_bias=True)(layer_h3) # LSTM层
|
||||
layer_h5 = Dropout(0.2)(layer_h4) # 随机中断部分神经网络连接,防止过拟合
|
||||
layer_h6 = Dense(1279, activation="softmax")(layer_h5) # 全连接层
|
||||
layer_h6 = Dense(self.MS_OUTPUT_SIZE, activation="softmax")(layer_h5) # 全连接层
|
||||
|
||||
#labels = Input(name='the_labels', shape=[60], dtype='float32')
|
||||
layer_out = Lambda(ctc_lambda_func,output_shape=(1279,), name='ctc')(layer_h6) # CTC
|
||||
layer_out = Lambda(ctc_lambda_func,output_shape=(self.MS_OUTPUT_SIZE, ), name='ctc')(layer_h6) # CTC
|
||||
_model = Model(inputs = layer_input, outputs = layer_out)
|
||||
|
||||
#_model = Sequential()
|
||||
|
||||
#_model.add(Conv1D(256, 5,input_shape=(1500,39), use_bias=True, padding="valid"))
|
||||
#_model.add(MaxPooling1D(pool_size=2, strides=None, padding="valid"))
|
||||
#_model.add(Dropout(0.3)) # 随机中断部分神经网络连接
|
||||
|
||||
#_model.add(LSTM(256, activation='relu', use_bias=True))
|
||||
#_model.add(Dropout(0.3)) # 随机中断部分神经网络连接
|
||||
|
||||
#_model.add(Dense(1279, activation="softmax"))
|
||||
##_model.add(Lambda(ctc_lambda_func,output_shape=(1,),name='ctc'))
|
||||
|
||||
#_model.compile(optimizer="sgd", loss='categorical_crossentropy',metrics=["accuracy"])
|
||||
_model.compile(optimizer="sgd", loss='ctc',metrics=["accuracy"])
|
||||
|
@ -68,7 +60,7 @@ class ModelSpeech(): # 语音模型类
|
|||
#labels, y_pred, input_length, label_length = args
|
||||
y_pred = args[:,2:,:]
|
||||
#y_pred = y_pred[:, 2:, :]
|
||||
return K.ctc_decode(y_pred,1279)
|
||||
return K.ctc_decode(y_pred,self.MS_OUTPUT_SIZE)
|
||||
#return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
|
||||
|
||||
def TrainModel(self,datapath,epoch = 2,save_step=1000,filename='model_speech/LSTM_CNN_model'):
|
||||
|
@ -84,11 +76,14 @@ class ModelSpeech(): # 语音模型类
|
|||
data.LoadDataList('train')
|
||||
num_data=DataSpeech.GetDataNum() # 获取数据的数量
|
||||
for epoch in range(epoch): # 迭代轮数
|
||||
print('[running] train epoch %d .' % epoch)
|
||||
n_step = 0 # 迭代数据数
|
||||
while True:
|
||||
try:
|
||||
print('[message] epoch %d . Have train datas %d+'%(epoch, n_step*save_step))
|
||||
# data_genetator是一个生成器函数
|
||||
self._model.fit_generator(data.data_genetator, save_step, nb_worker=2)
|
||||
yielddatas = data.data_genetator(self.BATCH_SIZE)
|
||||
self._model.fit_generator(yielddatas, save_step, nb_worker=2)
|
||||
n_step += 1
|
||||
except StopIteration:
|
||||
print('[error] generator error. please check data format.')
|
||||
|
@ -109,17 +104,28 @@ class ModelSpeech(): # 语音模型类
|
|||
'''
|
||||
self._model.save_weights(filename+comment+'.model')
|
||||
|
||||
def TestModel(self):
|
||||
def TestModel(self, datapath, str_dataset='dev'):
|
||||
'''
|
||||
测试检验模型效果
|
||||
'''
|
||||
pass
|
||||
data=DataSpeech(datapath)
|
||||
data.LoadDataList(str_dataset)
|
||||
num_data = DataSpeech.GetDataNum() # 获取数据的数量
|
||||
try:
|
||||
gen = data.data_genetator(num_data)
|
||||
for i in range(1):
|
||||
X, y = gen
|
||||
r = self._model.test_on_batch(X, y)
|
||||
print(r)
|
||||
except StopIteration:
|
||||
print('[Error] Model Test Error. please check data format.')
|
||||
|
||||
def Predict(self,x):
|
||||
'''
|
||||
预测结果
|
||||
'''
|
||||
r = predict_on_batch(x)
|
||||
r = self._model.predict_on_batch(x)
|
||||
print(r)
|
||||
return r
|
||||
pass
|
||||
|
||||
|
@ -132,5 +138,8 @@ class ModelSpeech(): # 语音模型类
|
|||
|
||||
|
||||
if(__name__=='__main__'):
|
||||
pass
|
||||
datapath = 'E:\\语音数据集'
|
||||
ms = ModelSpeech()
|
||||
ms.TrainModel(datapath)
|
||||
ms.TestModel(datapath)
|
||||
|
||||
|
|
10
readdata.py
10
readdata.py
|
@ -173,16 +173,6 @@ class DataSpeech():
|
|||
return v
|
||||
|
||||
if(__name__=='__main__'):
|
||||
#wave_data, fs = read_wav_data("general_function\\A2_0.wav")
|
||||
#print(wave_data)
|
||||
#(fs,wave_data)=wav.read('E:\\国创项目工程\代码\\ASRT_SpeechRecognition\\general_function\\A2_0.wav')
|
||||
#wav_show(wave_data[0],fs)
|
||||
#mfcc_feat = mfcc(wave_data[0],fs) # 计算MFCC特征
|
||||
#print(mfcc_feat[0:3,:])
|
||||
#d_mfcc_feat_1 = delta(mfcc_feat, 2)
|
||||
#print(d_mfcc_feat_1[0,:])
|
||||
#d_mfcc_feat_2 = delta(d_mfcc_feat_1, 2)
|
||||
#print(d_mfcc_feat_2[0,:])
|
||||
#path='E:\\语音数据集'
|
||||
#l=DataSpeech(path)
|
||||
#l.LoadDataList('train')
|
||||
|
|
Loading…
Reference in New Issue