def split_data(self, audiopath, labelpath):
"""
音频数据、标注地址分割
:param audiopath: 音频路径
:param labelpath: 标注路径
:return:
"""
generator = AcousticModel.init_audio(audiopath, labelpath) # 音频和标注的迭代器
file_count = generator.__next__() # 文件总数
chunk = file_count // self.num
for job_id in range(self.num - 1):
path_file = PARAMETERS_FILE_PATH + '/%s/pathInfo_%d.csv' % (
self.kwargs['unit_type'], job_id)
'''文件已存在'''
if os.path.exists(path_file):
continue
with open(path_file, 'w') as f:
count = 0
while count < chunk:
count += 1
path = generator.__next__()
f.write(path[0]) # 写入音频路径
f.write(path[1]) # 写入标注路径
'''写入最后一个路径列表'''
path_file = PARAMETERS_FILE_PATH + '/%s/pathInfo_%d.csv' % (
self.kwargs['unit_type'], (self.num - 1))
with open(path_file, 'w') as f:
for path in generator:
f.write(path[0]) # 写入音频路径
f.write(path[1]) # 写入标注路径
def main(record_seconds, n=2, savepath=path_record):
"""执行整个过程"""
global state_num, mfcc, unit_time, lexicon, unit, ngram
'''初始化声学模型'''
am = AcousticModel(mix_level=4)
am.initialize_unit(unit_type='XIF_tone')
am.init_parameter()
unit = am.unit
unit_time = dict([[key, -1] for key in list(am.unit.keys())])
state_num = am.statenum # 设置HMM状态数
'''初始化语言模型'''
for i in range(n):
ngram_ = Ngram(n=i + 1)
ngram_.init_gram()
ngram.append(ngram_)
'''初始化发音词典'''
p_lexicon = PronunciationLexicon()
p_lexicon.init_lexicon()
lexicon = p_lexicon.lexicon
'''获取MFCC帧'''
mfcc = record_audio(record_seconds, savepath)
'''生成首词模型'''
generate_first_word(am)
a = time.time()
'''执行令牌传递算法'''
token_passing()
b = time.time()
transfer()
print('耗时:', b - a)
def parallel_data(self, env, mode=1, init=True):
"""
并行处理数据
:param env: 用于标识机器的环境变量名
:param mode: 训练方案
:param init: 是否初始化
:return:
"""
'''获取作业ID'''
try:
job_id = os.environ[env]
except JobIDExistError:
raise JobIDExistError(self.log)
amodel = AcousticModel(self.log,
self.kwargs['unit_type'],
processes=self.kwargs['processes'],
job_id=job_id,
console=self.console,
state_num=self.kwargs['state_num'],
mix_level=self.kwargs['mix_level'],
delta_1=self.kwargs['delta_1'],
delta_2=self.kwargs['delta_2'])
amodel.load_unit()
amodel.process_data(mode=mode,
load_line=self.kwargs['load_line'],
init=init,
proportion=self.kwargs['proportion'],
step=self.kwargs['step'],
differentiation=self.kwargs['differentiation'],
coefficient=self.kwargs['coefficient'])
def split_unit(self):
"""
基元分割
:return: 各机器的训练清单
"""
amodel = AcousticModel(self.log,
self.kwargs['unit_type'],
processes=self.kwargs['processes'],
console=self.console,
state_num=self.kwargs['state_num'],
mix_level=self.kwargs['mix_level'],
delta_1=self.kwargs['delta_1'],
delta_2=self.kwargs['delta_2'])
amodel.load_unit()
units = amodel.loaded_units
chunk = len(units) // self.num
traininfo_path = PARAMETERS_FILE_PATH + '/%s' % self.kwargs['unit_type']
for job_id in range(self.num - 1):
trainInfo_file = traininfo_path + '/trainInfo_%d.csv' % job_id
'''删除现存trainInfo文件'''
if os.path.exists(trainInfo_file):
os.remove(trainInfo_file)
units_slice = set(units).difference(
set(units[job_id * chunk:(job_id + 1) * chunk]))
for unit in units_slice:
trainInfo = open(trainInfo_file, 'a+')
trainInfo.writelines('%s\n' % unit)
'''最后一个机器的训练清单'''
units_slice = set(units).difference(set(units[(self.num - 1) *
chunk:]))
trainInfo_file = traininfo_path + '/trainInfo_%d.csv' % (self.num - 1)
'''删除现存trainInfo文件'''
if os.path.exists(trainInfo_file):
os.remove(trainInfo_file)
for unit in units_slice:
trainInfo = open(
traininfo_path + '/trainInfo_%d.csv' % (self.num - 1), 'a+')
trainInfo.writelines('%s\n' % unit)
def parallel_train(self,
env,
init=True,
show_q=False,
show_a=False,
c_covariance=1e-3):
"""
并行训练
:param env: 用于标识机器的环境变量名
:param init: 是否初始化
:param show_q: 显示HMM/GMM当前似然度
:param show_a: 显示HMM重估后状态转移矩阵
:param c_covariance: 修正数值，纠正GMM中的Singular Matrix
:return:
"""
'''获取作业ID'''
try:
job_id = os.environ[env]
except JobIDExistError:
raise JobIDExistError(self.log)
'''从配置项中获取协方差修正数值，获取失败则在参数中获取，默认为1e-3'''
try:
c_covariance_ = os.environ['c_covariance']
except KeyError:
c_covariance_ = c_covariance
amodel = AcousticModel(self.log,
self.kwargs['unit_type'],
processes=self.kwargs['processes'],
job_id=job_id,
console=self.console,
state_num=self.kwargs['state_num'],
mix_level=self.kwargs['mix_level'],
delta_1=self.kwargs['delta_1'],
delta_2=self.kwargs['delta_2'])
amodel.load_unit()
amodel.training(init=init,
show_q=show_q,
show_a=show_a,
c_covariance=c_covariance_)
def parallel_data(self, env, init=True):
"""
并行处理数据
:param env: 用于标识机器的环境变量名
:param init: 是否初始化
:return:
"""
'''获取作业ID'''
try:
job_id = os.environ[env]
except JobIDExistError:
raise JobIDExistError(self.log)
amodel = AcousticModel(self.log,
self.kwargs['unit_type'],
processes=self.kwargs['processes'],
job_id=job_id,
console=self.console,
state_num=self.kwargs['state_num'],
mix_level=self.kwargs['mix_level'],
delta_1=self.kwargs['delta_1'],
delta_2=self.kwargs['delta_2'])
amodel.load_unit()
amodel.split_data(load_line=self.kwargs['load_line'], init=init)
def parallel_train(self,
env,
mode=1,
init=True,
show_q=False,
show_a=False):
"""
并行训练
:param env: 用于标识机器的环境变量名
:param mode: 训练方案
:param init: 是否初始化
:param show_q: 显示HMM/GMM当前似然度
:param show_a: 显示HMM重估后状态转移矩阵
:return:
"""
'''获取作业ID'''
try:
job_id = os.environ[env]
except JobIDExistError:
raise JobIDExistError(self.log)
amodel = AcousticModel(self.log,
self.kwargs['unit_type'],
processes=self.kwargs['processes'],
job_id=job_id,
console=self.console,
state_num=self.kwargs['state_num'],
mix_level=self.kwargs['mix_level'],
delta_1=self.kwargs['delta_1'],
delta_2=self.kwargs['delta_2'])
amodel.load_unit()
amodel.training(mode=mode,
init=init,
show_q=show_q,
show_a=show_a,
load_line=self.kwargs['load_line'],
c_covariance=self.c_covariance)