def load_and_proc_feats_segm(feats_dir,
files,
frame_step,
min_len,
max_len,
floatX='float32'):
feats_o = []
for f in files:
n_avl_samp = pytel.htk.readhtk_header(feats_dir + '/' + f + '.fea')[0]
n_sel_samp = np.random.randint(min_len, max_len)
if (n_avl_samp <= n_sel_samp):
feats_o.append(readhtk(feats_dir + '/' + f + '.fea'))
else:
# Randomly select the starting point
start = np.random.randint(0, n_avl_samp - n_sel_samp)
end = start + n_sel_samp
feats_o.append(
readhtk_segment(feats_dir + '/' + f + '.fea', start, end))
idx = np.cumsum(
np.vstack([len(feats_o[i]) for i in range(0, len(feats_o))]))
idx = np.insert(idx, 0, 0)
# Need these as arrays
feats_o = np.vstack(feats_o[i] for i in range(0, len(feats_o)))
return [feats_o, idx.astype(floatX)]
def load_jhu_feat_segm(feats_dir,
files,
min_len,
max_len,
floatX='float32',
rng=np.random,
start_from_zero=False):
feats = []
for f in files:
n_avl_samp = pytel.htk.readhtk_header(feats_dir + '/' + f + '.fea')[0]
n_sel_samp = rng.randint(min_len, max_len) # +1)
if (n_avl_samp <= n_sel_samp):
feats.append(readhtk(feats_dir + '/' + f + '.fea'))
else:
if start_from_zero:
start = 0
else:
start = rng.randint(0, n_avl_samp - n_sel_samp + 1)
end = start + n_sel_samp
feats.append(
readhtk_segment(feats_dir + '/' + f + '.fea', start, end))
idx = np.cumsum(np.vstack([len(feats[i]) for i in range(0, len(feats))]))
idx = np.insert(idx, 0, 0)
return np.vstack(feats)[np.newaxis, :, :], idx
def load_jhu_feat_and_vad(feats_dir, vad_dir, files, floatX='float32'):
feats = np.vstack([readhtk(feats_dir + '/' + f + '.fea') for f in files])
vad = [np.genfromtxt(vad_dir + '/' + f + '.vad', dtype=int) for f in files]
idx = np.cumsum(np.vstack([len(vad[i]) for i in range(0, len(vad))]))
idx = np.insert(idx, 0, 0)
return feats, idx, vad
def load_jhu_feat_lab_vad(feats_dir,
lab_dir,
vad_dir,
files,
max_frame=-1,
feat_is_segm=False,
floatX='float32'):
if (max_frame == -1):
feats = [readhtk(feats_dir + '/' + f + '.fea') for f in files]
else:
n_avl_samp = [
pytel.htk.readhtk_header(feats_dir + '/' + f + '.fea')[0]
for f in files
]
feats = [
readhtk_segment(feats_dir + '/' + f + '.fea',
start=0,
end=np.min((max_frame, n_avl_samp[i])))
for i, f in enumerate(files)
]
new_files = [fix_names_2(f) for f in files]
lab = []
for i, nf in enumerate(new_files):
if feat_is_segm:
f, start, end = grep_times(nf)
else:
f = nf
start = 0
end = feats[i].shape[0]
l = np.genfromtxt(lab_dir + f + ".hp")
if (vad_dir != None):
v = pytel.htk.read_lab_to_bool_vec(vad_dir + '/' + f + '.vad',
length=len(l),
true_label='speech')
l = l[v]
l = l[start:end]
if (len(l) != feats[i].shape[0]):
log.warning(
"WARNING: The length of lab is %d but lenght of features is %d. Is VAD missing on lab? "
% (len(l), feats[i].shape[0]))
lab.append(l[:, np.newaxis])
idx = np.cumsum(
np.vstack([len(feats[i]) for i in range(0, len(feats))]))
idx = np.insert(idx, 0, 0)
#print feats
print(lab[2].shape)
return np.vstack(feats)[np.newaxis, :, :], idx, np.vstack(lab)[
np.newaxis, :, :]
def load_and_proc_feats_expand(vad_dir,
feats_dir,
files,
frame_step=1,
floatX='float32',
max_length=None):
feats_o = [
readhtk(feats_dir + '/' + f + '.fea')[::frame_step, :] for f in files
]
feats = [
nn_def_anya.preprocess_nn_input_MFCC(
readhtk(feats_dir + '/' + f + '.fea')[0:max_length:frame_step, :])
for f in files
]
vad = [
pytel.htk.read_lab_to_bool_vec(vad_dir + '/' + files[i] + '.lab.gz',
true_label='sp',
length=-len(feats_o[i]))[15:-15]
for i in range(0, len(feats))
]
#vad = [pytel.htk.read_lab_to_bool_vec(vad_dir + '/' + files[i]+'.lab.gz', true_label='sp' )[15:-15]
# for i in range(0, len(feats)) ]
for i in range(0, len(feats)):
assert (len(vad) == feats[i].shape[0])
# Now apply the VAD
feats_o = np.vstack(feats_o[i][15:-15][vad[i]]
for i in range(0, len(feats)))
feats = np.vstack(feats[i][vad[i]] for i in range(0, len(feats)))
idx = np.cumsum(np.vstack([np.sum(vad[i]) for i in range(0, len(vad))]))
idx = np.insert(idx, 0, 0)
return [feats, feats_o, idx.astype(floatX)]
def load_feats_segm_vad_time_comb(vad_dir,
feats_dir,
files,
frame_step,
min_len,
max_len,
floatX='float32'):
feats_o = []
vad = []
for f in files:
n_avl_samp_tot = pytel.htk.readhtk_header(feats_dir + '/' + f +
'.fea')[0]
vad_tmp = pytel.htk.read_lab_to_bool_vec(vad_dir + '/' + f + '.lab.gz',
true_label='sp',
length=-n_avl_samp_tot)
assert (len(vad_tmp) == n_avl_samp_tot)
n_avl_samp = np.sum(vad_tmp) # The speech time
n_sel_samp = np.random.randint(min_len, max_len)
if (n_avl_samp <= n_sel_samp):
feats_o.append(readhtk(feats_dir + '/' + f + '.fea'))
vad.append(vad_tmp)
else:
# Randomly select the starting point
start = np.random.randint(0, n_avl_samp - n_sel_samp)
end = start + n_sel_samp
# Need to find the start and end index
idx = np.cumsum(vad_tmp)
start = int(np.where(np.cumsum(vad_tmp) == start)[0])
end = int(np.where(np.cumsum(vad_tmp) == end)[0])
feats_o.append(
readhtk_segment(feats_dir + '/' + f + '.fea', start, end))
vad.append(vad_tmp[start:end])
idx = np.cumsum(np.vstack([len(feats_o[i]) for i in range(0, len(vad))]))
idx = np.insert(idx, 0, 0)
# Need these as arrays
feats_o = np.vstack(feats_o[i] for i in range(0, len(feats_o)))
vad = np.hstack(vad[i] for i in range(0, len(vad)))
return [
np.hstack((feats_o, vad[:, np.newaxis].astype(floatX))),
idx.astype(floatX)
]
def load_jhu_feat(feats_dir, files, floatX='float32', max_frame=-1):
if (max_frame == -1):
feats = [readhtk(feats_dir + '/' + f + '.fea') for f in files]
else:
n_avl_samp = [
pytel.htk.readhtk_header(feats_dir + '/' + f + '.fea')[0]
for f in files
]
feats = [
readhtk_segment(feats_dir + '/' + f + '.fea',
start=0,
end=np.min((max_frame, n_avl_samp[i])))
for i, f in enumerate(files)
]
idx = np.cumsum(np.vstack([len(feats[i]) for i in range(0, len(feats))]))
idx = np.insert(idx, 0, 0)
return np.vstack(feats)[np.newaxis, :, :], idx
def load_and_proc_feats(feats_dir,
files,
frame_step=1,
floatX='float32',
max_length=None):
feats = []
idx = [0]
for f in files:
f_path = feats_dir + '/' + f + '.fea'
if (os.path.isfile(f_path)):
tmp_feats = readhtk(feats_dir + '/' + f +
'.fea')[0:max_length:frame_step, :]
idx.append(tmp_feats.shape[0])
feats.append(tmp_feats)
else:
idx.append(0)
idx = np.cumsum(np.vstack(idx))
feats = np.concatenate(feats)
return [feats.astype(floatX), idx.astype(floatX)]
def load_feats_comb(vad_dir, feats_dir, files, frame_step=1, floatX='float32'):
feats_o = [
readhtk(feats_dir + '/' + f + '.fea')[::frame_step, :] for f in files
]
vad = [
pytel.htk.read_lab_to_bool_vec(vad_dir + '/' + files[i] + '.lab.gz',
true_label='sp',
length=-len(feats_o[i]))
for i in range(0, len(feats_o))
]
idx = np.cumsum(np.vstack([len(feats_o[i]) for i in range(0, len(vad))]))
idx = np.insert(idx, 0, 0)
# Need these as arrays
feats_o = np.vstack(feats_o[i] for i in range(0, len(feats_o)))
vad = np.hstack(vad[i] for i in range(0, len(vad)))
#return [ np.hstack( ( feats_o, vad[:,np.newaxis].astype(T.config.floatX) ) ), idx.astype('float32')
return [
np.hstack((feats_o, vad[:, np.newaxis].astype(floatX))),
idx.astype(floatX)
]
def load_feats(vad_dir, feats_dir, files, frame_step=1, floatX='float32'):
feats_o = [
readhtk(feats_dir + '/' + f + '.fea')[::frame_step, :] for f in files
]
#feats = [ nn_def_anya.preprocess_nn_input_MFCC( readhtk(feats_dir + '/'+ f + '.fea')[::frame_step,:] ) for f in files]
vad = [
pytel.htk.read_lab_to_bool_vec(vad_dir + '/' + files[i] + '.lab.gz',
true_label='sp',
length=-len(feats_o[i]))
for i in range(0, len(feats_o))
]
for i in range(0, len(feats_o)):
assert (len(vad) == feats_o[i].shape[0])
idx = np.cumsum(np.vstack([len(feats_o[i]) for i in range(0, len(vad))]))
idx = np.insert(idx, 0, 0)
# Need these as arrays
feats_o = np.vstack(feats_o[i] for i in range(0, len(feats_o)))
vad = np.hstack(vad[i] for i in range(0, len(vad)))
return [feats_o, vad, idx.astype(floatX)]
def load_feats(feat_dir, files, max_length, frame_step):
d = [readhtk(feat_dir + '/'+ f + '.fea')[::frame_step,:] for f in files]
return [pad_sequences(d, max_length, dtype=float)]
