def prepare_training_data(chime_data_dir, dest_dir, suffix_id):
for stage in ['tr', 'dt']:
fpath, flist = gen_flist_simu(chime_data_dir, stage, suffix_id)
export_flist = list()
mkdir_p(os.path.join(dest_dir, stage))
for f in tqdm.tqdm(flist, desc='Generating data for {}'.format(fpath)):
clean_audio = get_audio_data(f, '_clean')
noise_audio = get_audio_data(f, '_noise')
X = stft(clean_audio, time_dim=1).transpose((1, 0, 2))
N = stft(noise_audio, time_dim=1).transpose((1, 0, 2))
IBM_X, IBM_N = estimate_IBM(X, N)
Y_abs = np.abs(X + N)
export_dict = {
'IBM_X': IBM_X.astype(np.float32),
'IBM_N': IBM_N.astype(np.float32),
'Y_abs': Y_abs.astype(np.float32)
}
export_name = os.path.join(dest_dir, stage, f.split('/')[-1])
with open(export_name, 'wb') as fid:
pickle.dump(export_dict, fid)
export_flist.append(os.path.join(stage, f.split('/')[-1]))
with open(
os.path.join(dest_dir,
'flist_{}_{}.json'.format(stage, suffix_id)),
'w') as fid:
json.dump(export_flist, fid, indent=4)
def prepare_training_data(chime_data_dir, dest_dir):
for stage in ['tr', 'dt']:
flist = gen_flist_simu(chime_data_dir, stage, ext=True)
export_flist = list()
# mkdir_p(os.path.join(dest_dir, stage))
for idx in range(len(flist)):
f = flist[idx]
clean_audio = get_audio_data(f, '.Clean')
noise_audio = get_audio_data(f, '.Noise')
X = stft(clean_audio, time_dim=1, size=256, shift=128).transpose(
(1, 0, 2))
N = stft(noise_audio, time_dim=1, size=256, shift=128).transpose(
(1, 0, 2))
IBM_X, IBM_N = estimate_IBM(X, N)
Y_abs = np.abs(X + N)
export_dict = {
'IBM_X': IBM_X.astype(np.float32),
'IBM_N': IBM_N.astype(np.float32),
'Y_abs': Y_abs.astype(np.float32)
}
export_name = os.path.join(dest_dir, stage, f.split('/')[-1])
with open(export_name, 'wb') as fid:
pickle.dump(export_dict, fid)
export_flist.append(os.path.join(stage, f.split('/')[-1]))
with open(os.path.join(dest_dir, 'flist_{}.json'.format(stage)),
'w') as fid:
json.dump(export_flist, fid, indent=4)
def prepare_clean_training_data(chime_data_dir, dest_dir):
start = 0
# print("sdsd")
for stage in ['tr', 'dt']:
reset_counter = 0
flist = gen_flist_simu(chime_data_dir, stage, ext=True)
# print(flist)
export_flist = list()
mkdir_p(os.path.join(dest_dir, stage))
clean_data = audioread('/media/hipo/Mega Store/Dataset/single file/Chinese_tai_clean.wav')
print("clean_data size:", clean_data.shape[0])
for f in tqdm.tqdm(flist, desc='Generating data for {}'.format(stage)):
# clean_audio = get_audio_data(f, '.Clean')
noise_audio = get_audio_data(f, '.Noise')
# print(chime_data_dir)
chime_size = audioread('{}.CH{}{}.Noise.wav'.format(f, 1, ''))
clean_files = list()
end = chime_size.shape[0] + start
if end > clean_data.shape[0]:
print("reset counter: ", reset_counter + 1)
start = 0
end = chime_size.shape[0] + start
for i in range(1, 7):
y = clean_data[start:end]
start = end
clean_files.append(y[None, :])
clean_files = np.concatenate(clean_files, axis=0)
clean_files = clean_files.astype(np.float32)
clean_audio = clean_files
X = stft(clean_audio, time_dim=1).transpose((1, 0, 2))
N = stft(noise_audio, time_dim=1).transpose((1, 0, 2))
IBM_X, IBM_N = estimate_IBM(X, N)
Y_abs = np.abs(X + N)
export_dict = {
'IBM_X': IBM_X.astype(np.float32),
'IBM_N': IBM_N.astype(np.float32),
'Y_abs': Y_abs.astype(np.float32)
}
export_name = os.path.join(dest_dir, stage, f.split('/')[-1])
with open(export_name, 'wb') as fid:
pickle.dump(export_dict, fid)
export_flist.append(os.path.join(stage, f.split('/')[-1]))
with open(os.path.join(dest_dir, 'flist_{}.json'.format(stage)),
'w') as fid:
json.dump(export_flist, fid, indent=4)
def prepare_training_data(chime_data_dir, dest_dir):
for stage in ['tr', 'dt']:
flist = gen_flist_simu(chime_data_dir, stage, ext=True)
export_flist = list()
mkdir_p(os.path.join(dest_dir, stage))
for f in tqdm.tqdm(flist, desc='Generating data for {}'.format(stage)):
clean_audio = get_audio_data(f, '.Clean')
noise_audio = get_audio_data(f, '.Noise')
X = stft(clean_audio, time_dim=1).transpose((1, 0, 2))
N = stft(noise_audio, time_dim=1).transpose((1, 0, 2))
IBM_X, IBM_N = estimate_IBM(X, N)
Y_abs = np.abs(X + N)
export_dict = {
'IBM_X': IBM_X.astype(np.float32),
'IBM_N': IBM_N.astype(np.float32),
'Y_abs': Y_abs.astype(np.float32)
}
export_name = os.path.join(dest_dir, stage, f.split('/')[-1])
with open(export_name, 'wb') as fid:
pickle.dump(export_dict, fid)
export_flist.append(os.path.join(stage, f.split('/')[-1]))
with open(os.path.join(dest_dir, 'flist_{}.json'.format(stage)),
'w') as fid:
json.dump(export_flist, fid, indent=4)
def prepare_other_training_data(train_dir, dest_dir):
start = 0
chime_data_dir = os.path.join(train_dir[:-1], 'tr')
print(chime_data_dir)
for stage in ['tr', 'dt']:
if stage is 'dt':
chime_data_dir = os.path.join(train_dir[:-1], 'dt')
print(chime_data_dir)
reset_counter = 0
# flist = gen_flist_simu(chime_data_dir, stage, ext=True)
flist = [f for f in listdir(chime_data_dir) if isfile(join(chime_data_dir, f))]
# print(flist)
export_flist = list()
mkdir_p(os.path.join(dest_dir, stage))
noise_data = audioread('/media/hipo/lento/Dataset/single file/noise_files/all_noise.wav')
print("noise_data size:", noise_data.shape[0])
for f in tqdm.tqdm(flist, desc='Generating data for {}'.format(stage)):
# clean_audio = get_audio_data(f)
path = os.path.join(chime_data_dir, f)
clean_audio = get_audio_single(path)
# clean_audioa = audioread(path)
# clean_audiob = audioread(path)
# multi_track = list()
# multi_track.append(clean_audioa[None, :])
# multi_track.append(clean_audiob[None, :])
# multi_track = np.concatenate(multi_track, axis=0)
# multi_track = multi_track.astype(np.float32)
# print(multi_track.shape)
chime_size = audioread(path)
noise_files = list()
end = chime_size.shape[0] + start
if end > noise_data.shape[0]:
print("reset counter: ", reset_counter + 1)
start = 0
end = chime_size.shape[0] + start
for i in range(1, 2):
y = noise_data[start:end]
start = end
noise_files.append(y[None, :])
noise_files = np.concatenate(noise_files, axis=0)
noise_files = noise_files.astype(np.float32)
noise_audio = noise_files
# print("speech size: ", multi_track.shape, "noise size: ", noise_audio.shape)
X = stft(clean_audio, time_dim=1).transpose((1, 0, 2))
N = stft(noise_audio, time_dim=1).transpose((1, 0, 2))
IBM_X, IBM_N = estimate_IBM(X, N)
Y_abs = np.abs(X + N)
export_dict = {
'IBM_X': IBM_X.astype(np.float32),
'IBM_N': IBM_N.astype(np.float32),
'Y_abs': Y_abs.astype(np.float32)
}
export_name = os.path.join(dest_dir, stage, f.split('/')[-1])
with open(export_name, 'wb') as fid:
pickle.dump(export_dict, fid)
export_flist.append(os.path.join(stage, f.split('/')[-1]))
with open(os.path.join(dest_dir, 'flist_{}.json'.format(stage)),
'w') as fid:
json.dump(export_flist, fid, indent=4)
def get_batch(self, uttid, channel=0, divisor=16, max_length=np.inf):
""" Load a batch of data from files """
batch = {}
if 'numpy' in self.flists:
fname = self.flists['numpy']['data'][uttid]
feats = load_arrays_from_numpy(self.base_dir, fname)
else:
feats = {}
start_idx = None
for name in ['clean', 'noisy', 'noise'] & self.flists.keys():
fname = self.flists[name]['data'][uttid]
if self.flists[name]['type'] == 'json':
feats[name] = load_arrays_from_wav(self.base_dir,
fname,
idx=channel)
feats[name] = np.expand_dims(feats[name], axis=0)
elif self.flists[name]['type'] == 'scp':
feats[name] = load_arrays_from_scp(
self.base_dir, fname, remove_deltas=name == 'noisy')
else:
raise ValueError("Type must be one of 'json', 'scp'")
if 'noise' in feats:
feats['noisy'] = feats['clean'] + feats['noise']
del feats['noise']
# Reshape features
for name in ['clean', 'noisy'] & feats.keys():
if feats[name].dtype == 'complex64':
feats[name] = np.abs(feats[name])
if self.logify:
feats[name] = np.log(feats[name] + 0.01)
if 'senone' in self.flists:
feats['senone'] = self.flists['senone']['data'][uttid]
feats['senone'] = np.expand_dims(feats['senone'], axis=0)
if 'senone' in feats:
feats['senone'] = np.expand_dims(feats['senone'], axis=0)
if feats['senone'].shape[2] < feats['clean'].shape[2]:
padding = [
(0, 0), (0, 0),
(0, feats['clean'].shape[2] - feats['senone'].shape[2])
]
feats['senone'] = np.pad(feats['senone'], padding, 'edge')
start = -1
for name in ['clean', 'noisy', 'noise', 'senone'] & feats.keys():
if feats[name].shape[2] > max_length:
if start == -1:
start = np.random.randint(feats[name].shape[2] -
max_length)
feats[name] = feats[name][:, :, start:start + max_length]
feats['frames'] = feats[name].shape[-2]
#feats = shrink_to_min(feats, out_shape, max_length)
if self.compute_ibm:
if 'noise' in feats and 'clean' in feats:
feats['ibm_x'], feats['ibm_n'] = estimate_IBM(
feats['clean'], feats['noise'])
elif 'noisy' in feats and 'clean' in feats:
feats['ibm_x'], feats['ibm_n'] = estimate_IBM(
feats['clean'], feats['noisy'], -15, -15)
else:
raise ValueError(
"To compute IBM, clean and noise or noisy signals are required"
)
if self.compute_irm:
if 'noise' in feats and 'clean' in feats:
feats['irm'] = feats['clean'] / (feats['clean'] +
feats['noise'])
#feats['irm'][feats['irm'] > 1] = 1
elif 'noisy' in feats and 'clean' in feats:
if np.min(feats['clean']) < 0:
minimum = min(np.min(feats['clean']),
np.min(feats['noisy']))
feats['irm'] = (feats['clean'] - minimum +
1e-6) / (feats['noisy'] - minimum + 1e-6)
#feats['irm'] /= 1.3
else:
feats['irm'] = np.sqrt(feats['clean']) / np.sqrt(
feats['noisy'])
#feats['irm'] /= 2
feats['irm'][feats['irm'] > 1] = 1
else:
raise ValueError(
"To compute IRM, clean and noise or noisy signals are required"
)
#if 'noise' in feats:
# feats['noisy'] = feats['noise'] + feats['clean']
if 'trans' in self.flists:
indices = np.array(
[(0, i)
for i in range(len(self.flists['trans']['data'][uttid]))],
dtype=np.int32)
values = np.array(self.flists['trans']['data'][uttid],
dtype=np.int32)
shape = np.array((1, len(self.flists['trans']['data'][uttid])),
dtype=np.int32)
feats['trans'] = (indices, values, shape)
return feats