def feature_normalisation(self, in_file_list, out_file_list):
logger = logging.getLogger('feature_normalisation')
# self.feature_dimension = feature_dimension
try:
assert len(in_file_list) == len(out_file_list)
except AssertionError:
logger.critical('The input and output file numbers are not the same! %d vs %d' %(len(in_file_list), len(out_file_list)))
raise
if self.mean_vector == None:
self.mean_vector = self.compute_mean(in_file_list, 0, self.feature_dimension)
if self.std_vector == None:
self.std_vector = self.compute_std(in_file_list, self.mean_vector, 0, self.feature_dimension)
io_funcs = HTKFeat_read()
file_number = len(in_file_list)
for i in xrange(file_number):
features, current_frame_number = io_funcs.getall(in_file_list[i])
# print current_frame_number
# features = io_funcs.data
# current_frame_number = io_funcs.n_samples
mean_matrix = numpy.tile(self.mean_vector, (current_frame_number, 1))
std_matrix = numpy.tile(self.std_vector, (current_frame_number, 1))
norm_features = (features - mean_matrix) / std_matrix
htk_writer = HTKFeat_write(veclen=io_funcs.veclen, sampPeriod=io_funcs.sampPeriod, paramKind=9)
htk_writer.writeall(norm_features, out_file_list[i])
# htk_writter = HTK_Parm_IO(n_samples=io_funcs.n_samples, samp_period=io_funcs.samp_period, samp_size=io_funcs.samp_size, param_kind=io_funcs.param_kind, data=norm_features)
# htk_writter.write_htk(out_file_list[i])
return self.mean_vector, self.std_vector
def compute_mean(self, file_list, start_index, end_index):
logger = logging.getLogger('feature_normalisation')
local_feature_dimension = end_index - start_index
mean_vector = numpy.zeros((1, local_feature_dimension))
all_frame_number = 0
io_funcs = HTKFeat_read()
for file_name in file_list:
features, current_frame_number = io_funcs.getall(file_name)
# io_funcs = HTK_Parm_IO()
# io_funcs.read_htk(file_name)
# features = io_funcs.data
# current_frame_number = io_funcs.n_samples
mean_vector += numpy.reshape(numpy.sum(features[:, start_index:end_index], axis=0), (1, local_feature_dimension))
all_frame_number += current_frame_number
mean_vector /= float(all_frame_number)
# setting the print options in this way seems to break subsequent printing of numpy float32 types
# no idea what is going on - removed until this can be solved
# po=numpy.get_printoptions()
# numpy.set_printoptions(precision=2, threshold=20, linewidth=1000, edgeitems=4)
logger.info('computed mean vector of length %d :' % mean_vector.shape[1] )
logger.info(' mean: %s' % mean_vector)
# restore the print options
# numpy.set_printoptions(po)
self.mean_vector = mean_vector
return mean_vector
def compute_std(self, file_list, mean_vector, start_index, end_index):
logger = logging.getLogger('feature_normalisation')
local_feature_dimension = end_index - start_index
std_vector = numpy.zeros((1, self.feature_dimension))
all_frame_number = 0
io_funcs = HTKFeat_read()
for file_name in file_list:
features, current_frame_number = io_funcs.getall(file_name)
mean_matrix = numpy.tile(mean_vector, (current_frame_number, 1))
std_vector += numpy.reshape(numpy.sum((features[:, start_index:end_index] - mean_matrix) ** 2, axis=0), (1, local_feature_dimension))
all_frame_number += current_frame_number
std_vector /= float(all_frame_number)
std_vector = std_vector ** 0.5
# setting the print options in this way seems to break subsequent printing of numpy float32 types
# no idea what is going on - removed until this can be solved
# po=numpy.get_printoptions()
# numpy.set_printoptions(precision=2, threshold=20, linewidth=1000, edgeitems=4)
logger.info('computed std vector of length %d' % std_vector.shape[1] )
logger.info(' std: %s' % std_vector)
# restore the print options
# numpy.set_printoptions(po)
self.std_vector = std_vector
return std_vector
def compute_std(self, file_list, mean_vector, start_index, end_index):
logger = logging.getLogger('feature_normalisation')
local_feature_dimension = end_index - start_index
std_vector = numpy.zeros((1, self.feature_dimension))
all_frame_number = 0
io_funcs = HTKFeat_read()
for file_name in file_list:
features, current_frame_number = io_funcs.getall(file_name)
std_vector += numpy.reshape(
numpy.sum(
(features[:, start_index:end_index] - mean_vector)**2,
axis=0), (1, local_feature_dimension))
all_frame_number += current_frame_number
std_vector /= float(all_frame_number)
std_vector = std_vector**0.5
# setting the print options in this way seems to break subsequent printing of numpy float32 types
# no idea what is going on - removed until this can be solved
# po=numpy.get_printoptions()
# numpy.set_printoptions(precision=2, threshold=20, linewidth=1000, edgeitems=4)
logger.info('computed std vector of length %d' % std_vector.shape[1])
logger.info(' std: %s' % std_vector)
# restore the print options
# numpy.set_printoptions(po)
self.std_vector = std_vector
return std_vector
def prepare_training(self, file_id_list_name, wav_dir, lab_dir, work_dir,
multiple_speaker):
print('---preparing enverionment')
self.cfg_dir = os.path.join(work_dir, 'config')
self.model_dir = os.path.join(work_dir, 'model')
self.cur_dir = os.path.join(self.model_dir, 'hmm0')
if not os.path.exists(self.cfg_dir):
os.makedirs(self.cfg_dir)
if not os.path.exists(self.cur_dir):
os.makedirs(self.cur_dir)
self.phonemes = os.path.join(work_dir, 'mono_phone.list')
self.phoneme_map = os.path.join(work_dir, 'phoneme_map.dict')
# HMMs
self.proto = os.path.join(self.cfg_dir, 'proto')
# SCP files
self.copy_scp = os.path.join(self.cfg_dir, 'copy.scp')
self.test_scp = os.path.join(self.cfg_dir, 'test.scp')
self.train_scp = os.path.join(self.cfg_dir, 'train.scp')
# CFG
self.cfg = os.path.join(self.cfg_dir, 'cfg')
self.wav_dir = wav_dir
self.lab_dir = lab_dir
self.mfc_dir = os.path.join(work_dir, 'mfc')
if not os.path.exists(self.mfc_dir):
os.makedirs(self.mfc_dir)
self.mono_lab_dir = os.path.join(work_dir, 'mono_no_align')
if not os.path.exists(self.mono_lab_dir):
os.makedirs(self.mono_lab_dir)
file_id_list = self._read_file_list(file_id_list_name)
print('---checking data')
speaker_utt_dict = self._check_data(file_id_list, multiple_speaker)
print('---extracting features')
self._HCopy()
print(time.strftime("%c"))
print('---feature_normalisation')
io_funcs = HTKFeat_read()
htk_writer = HTKFeat_write(veclen=io_funcs.veclen,
sampPeriod=io_funcs.sampPeriod,
paramKind=9)
normaliser = Statis(feature_dimension=39,
read_func=io_funcs.getall,
writer_func=htk_writer.writeall)
for key_name in list(speaker_utt_dict.keys()):
normaliser.feature_normalisation(
speaker_utt_dict[key_name],
speaker_utt_dict[key_name]) ## save to itself
print(time.strftime("%c"))
print('---making proto')
self._make_proto()
def feature_normalisation(self, in_file_list, out_file_list):
logger = logging.getLogger('feature_normalisation')
# self.feature_dimension = feature_dimension
try:
assert len(in_file_list) == len(out_file_list)
except AssertionError:
logger.critical(
'The input and output file numbers are not the same! %d vs %d'
% (len(in_file_list), len(out_file_list)))
raise
# TODO: Compute them in one pass.
if self.mean_vector == None:
self.mean_vector = self.compute_mean(in_file_list, 0,
self.feature_dimension)
if self.std_vector == None:
self.std_vector = self.compute_std(in_file_list, self.mean_vector,
0, self.feature_dimension)
io_funcs = HTKFeat_read()
file_number = len(in_file_list)
for i in range(file_number):
features, current_frame_number = io_funcs.getall(in_file_list[i])
# print current_frame_number
# features = io_funcs.data
# current_frame_number = io_funcs.n_samples
# mean_matrix = numpy.tile(self.mean_vector, (current_frame_number, 1)) # Use numpy broadcasting instead.
# std_matrix = numpy.tile(self.std_vector, (current_frame_number, 1))
norm_features = (features - self.mean_vector) / self.std_vector
htk_writer = HTKFeat_write(veclen=io_funcs.veclen,
sampPeriod=io_funcs.sampPeriod,
paramKind=9)
htk_writer.writeall(norm_features, out_file_list[i])
# htk_writter = HTK_Parm_IO(n_samples=io_funcs.n_samples, samp_period=io_funcs.samp_period, samp_size=io_funcs.samp_size, param_kind=io_funcs.param_kind, data=norm_features)
# htk_writter.write_htk(out_file_list[i])
return self.mean_vector, self.std_vector
def get_data_4_predict(x_dir, y_dir, is_y=True):
if is_y:
x_tmp = []
y_tmp = []
f_tmp = []
for item in os.listdir(x_dir):
if item.endswith(".htk"):
# read the mfcc features
reader = HTKFeat_read(x_dir + item)
matrix = reader.getall()
x_tmp.append(matrix)
labels = np.loadtxt(y_dir + item.replace("_16.htk", '.txt'))
y_tmp.append(labels)
f_tmp.append([item, len(labels)])
x = np.array(x_tmp)
y = np.array(y_tmp)
f_names = np.array(f_tmp)
return x, y, f_names
else:
x_tmp = []
f_tmp = []
for item in os.listdir(x_dir):
if item.endswith(".htk"):
# read the mfcc features
reader = HTKFeat_read(x_dir + item)
matrix = reader.getall()
x_tmp.append(matrix)
f_tmp.append([item, len(matrix)])
x = np.array(x_tmp)
f_names = np.array(f_tmp)
return x, f_names
def compute_mean(self, file_list, start_index, end_index):
logger = logging.getLogger('feature_normalisation')
local_feature_dimension = end_index - start_index
mean_vector = numpy.zeros((1, local_feature_dimension))
all_frame_number = 0
io_funcs = HTKFeat_read()
for file_name in file_list:
features, current_frame_number = io_funcs.getall(file_name)
# io_funcs = HTK_Parm_IO()
# io_funcs.read_htk(file_name)
# features = io_funcs.data
# current_frame_number = io_funcs.n_samples
mean_vector += numpy.reshape(
numpy.sum(features[:, start_index:end_index], axis=0),
(1, local_feature_dimension))
all_frame_number += current_frame_number
mean_vector /= float(all_frame_number)
# setting the print options in this way seems to break subsequent printing of numpy float32 types
# no idea what is going on - removed until this can be solved
# po=numpy.get_printoptions()
# numpy.set_printoptions(precision=2, threshold=20, linewidth=1000, edgeitems=4)
logger.info('computed mean vector of length %d :' %
mean_vector.shape[1])
logger.info(' mean: %s' % mean_vector)
# restore the print options
# numpy.set_printoptions(po)
self.mean_vector = mean_vector
return mean_vector