def extract(self, url_list):
labels = [item[1] for item in url_list]
url_list = [item[0] for item in url_list]
self.logger.info("Start extract %d audio." % len(url_list))
param = zip(url_list, [self.sample_rate for i in range(len(url_list))],
[self.hop_length for i in range(len(url_list))],
[self.fix_len for i in range(len(url_list))])
ys = ops.multi_processing(self._extract_one, param, self.process_num)
if self.slides != [None, None] and self.slides != [0, 0]:
slides = [self.slides for i in range(len(ys))]
ys = ops.multi_processing(slide_windows, zip(ys, slides), self.process_num)
return ys, labels
def run(config):
reader = TFrecordClassBalanceGen(config, 'train')
train_urls = [os.path.join(config.save_path, 'url/train_tmp_0')]
for train_url in train_urls:
x, y = ext_fbank_feature(train_url, config)
x = ops.multi_processing(limit_len, x, config.n_threads, True)
reader.write(x, y)
del x, y
def ext_mfcc_feature(url_path, config):
"""This function is used for extract MFCC feature of a dataset.
Parameters
----------
url_path : ``str``
The path of the 'PATH' file.
config : ``config``
config of feature. (To decide if we need slide_window, and params of slide_window)
Returns
-------
fbank : ``list``
The feature array. each frame concat with the frame before and after it.
label : ``list``
The label of fbank feature.
"""
logger = logging.getLogger('data')
with open(url_path, 'r') as urls:
labels = []
url_list = []
for url in list(urls):
line, label = str(url).split(" ")
index = eval(str(label).split("\n")[0])
labels.append([index])
url_list.append(line)
logger.info("Extracting MFCC feature, utt_nums is %d" % len(url_list))
n_filt = [config.feature_dims for i in range(len(url_list))]
slides = [config.slides for i in range(len(url_list))]
if hasattr(config, 'min_db'):
dbs = [config.min_db for i in range(len(url_list))]
mfccs = ops.multi_processing(calc_mfcc,
zip(url_list, n_filt, slides, dbs),
config.n_threads)
else:
mfccs = ops.multi_processing(calc_mfcc, zip(url_list, n_filt, slides),
config.n_threads)
logger.info("Extracting MFCC feature succeed")
return mfccs, labels
def extract(self, url_list):
paired_data = self.url_pair(url_list, self.source_num)
self.logger.info("Start extract %d audio." % len(url_list))
param = zip(paired_data,
[self.sample_rate for i in range(len(url_list))],
[self.n_fft for i in range(len(url_list))],
[self.dims for i in range(len(url_list))],
[self.config.hop_length for i in range(url_list)],
[self.fix_len for i in range(len(url_list))])
spec_dic = ops.multi_processing(self._extract_one, param, self.process_num)
return spec_dic['data'], spec_dic['label']
def ext_fbank_feature(url_path, config):
"""This function is used for extract features of one dataset.
Parameters
----------
url_path : ``str``
The path of the 'PATH' file.
config : ``config``
config of feature. (Contain the parameters of slide_window and feature_dims)
Returns
-------
fbank : ``list``
The feature array. each frame concat with the frame before and after it.
label : ``list``
The label of fbank feature.
Notes
-----
Changeable concat size is in the todolist
"""
logger = logging.getLogger('data')
url_list = []
labels = []
with open(url_path, 'r') as urls:
for line in list(urls):
url, label = str(line).split(" ")
index = eval(str(label).split("\n")[0])
url_list.append(url)
labels.append([index])
logger.info("Extracting fbank feature, utt_nums is %d" % len(url_list))
if config.fix_len is not None:
max_len = get_max_audio_time(url_list)
max_lens = [max_len for i in range(len(url_list))]
else:
max_lens = [None for i in range(len(url_list))]
n_filt = [config.feature_dims for i in range(len(url_list))]
if config.slides is not None:
slide_l = [config.slides[0] for i in range(len(url_list))]
slide_r = [config.slides[1] for i in range(len(url_list))]
else:
slide_l = [None for i in range(len(url_list))]
slide_r = [None for i in range(len(url_list))]
fbanks = ops.multi_processing(
calc_fbank, zip(url_list, n_filt, slide_l, slide_r, max_lens),
config.n_threads)
logger.info("Extracting fbank feature succeed")
return fbanks, labels
def ext_fbank_feature(url_path, config):
"""This function is used for extract features of one dataset.
Parameters
----------
url_path : ``str``
The path of the 'PATH' file.
config : ``config``
config of feature. (To decide if we need slide_window, and params of slide_window)
Returns
-------
fbank : ``list``
The feature array. each frame concat with the frame before and after it.
label : ``list``
The label of fbank feature.
Notes
-----
Changeable concat size is in the todolist
"""
logger = logging.getLogger(config.model_name)
url_list = []
labels = []
with open(url_path, 'r') as urls:
for line in list(urls):
url, label = str(line).split(" ")
index = eval(str(label).split("\n")[0])
url_list.append(url)
labels.append([index])
logger.info("Extracting fbank feature, utt_nums is %d" % len(url_list))
n_filt = [config.feature_dims for i in range(len(url_list))]
slides = [config.slides for i in range(len(url_list))]
fbanks = ops.multi_processing(calc_fbank, zip(url_list, n_filt, slides),
config.n_threads)
logger.info("Extracting fbank feature succeed")
return fbanks, labels
def ext_spec_feature(url_path, config):
logger = logging.getLogger('data')
with open(url_path, 'r') as urls:
labels = []
url_list = []
for url in list(urls):
line, label = str(url).split(" ")
index = eval(str(label).split("\n")[0])
labels.append([index])
url_list.append(line)
logger.info("Extracting Spec feature, utt_nums is %d" % len(url_list))
NFFT = [config.NFFT for i in range(len(url_list))]
frame_size = [config.frame_size for i in range(len(url_list))]
(mag_spec,
phase_spec) = ops.multi_processing(get_stft,
zip(url_list, NFFT, frame_size),
config.n_threads)
logger.info("Extracting Spec feature succeed")
return mag_spec, phase_spec
if __name__ == '__main__':
config = TrainConfig('../config.json')
config.save_path = '.'
train_urls = [
'/home/data/speaker-recognition/url/train_1',
'/home/data/speaker-recognition/url/train_2',
'/home/data/speaker-recognition/url/train_3'
]
enroll_url = '/home/data/speaker-recognition/url/enroll'
test_url = '/home/data/speaker-recognition/url/test'
gen_train = TFrecordGen(config, 'Train.record')
for train_url in train_urls:
x, y = ext_fbank_feature(train_url, config)
x = ops.multi_processing(limit_len, x, config.n_threads, True)
gen_train.write(x, y)
del x, y
gen_enroll = TFrecordGen(config, 'Enroll.record')
x, y = ext_fbank_feature(enroll_url, config)
x = ops.multi_processing(limit_len, x, config.n_threads, True)
gen_enroll.write(x, y)
gen_test = TFrecordGen(config, 'Test.record')
x, y = ext_fbank_feature(test_url, config)
x = ops.multi_processing(limit_len, x, config.n_threads, True)
gen_test.write(x, y)
logger = logging.getLogger(config.model_name + '_train')
#logger.info("Feature proccessing done.")
