def acoustic_decomposition(in_file_list, dimension_all):
param_dim_dict = {'mgc': 180, 'vuv': 1, 'lf0': 3, 'bap': 3}
file_ext_dict = {'mgc': '.mgc', 'lf0': '.lf0', 'bap': '.bap'}
# load parameter variance
param_var_dir = \
'/home/brycezou/works/tuling_tts_train/egs/005_hdl_test/s1/experiments/ht-500-data/acoustic_model/data/var'
var_file_dict = {}
for feature_name in param_dim_dict.keys():
var_file_dict[feature_name] = os.path.join(param_var_dir, feature_name + '_' + str(param_dim_dict[feature_name]))
var_dict = {}
for feature_name in var_file_dict.keys():
var_value, _ = dp.load_binary_file_frame(var_file_dict[feature_name], 1)
var_value = np.reshape(var_value, (param_dim_dict[feature_name], 1))
var_dict[feature_name] = var_value
# parameter start index
dimension_index = 0
stream_start_index = {}
for feature_name in param_dim_dict.keys():
stream_start_index[feature_name] = dimension_index
dimension_index += param_dim_dict[feature_name]
wave_feature_types = ['mgc', 'lf0', 'bap']
inf_float = -1.0e+10
mlpg = MLParameterGeneration()
# one cmp file per loop
for file_name in in_file_list:
dir_name = os.path.dirname(file_name)
file_id = os.path.splitext(os.path.basename(file_name))[0]
# load cmp data from file
features_all, frame_number = dp.load_binary_file_frame(file_name, dimension_all)
# one type of features per loop
for feature_name in wave_feature_types:
curr_feature = features_all[:, stream_start_index[feature_name]: \
stream_start_index[feature_name]+param_dim_dict[feature_name]]
var = var_dict[feature_name]
var = np.transpose(np.tile(var, frame_number))
gen_features = mlpg.generation(curr_feature, var, param_dim_dict[feature_name]/3)
if feature_name in ['lf0', 'F0']:
if stream_start_index.has_key('vuv'):
vuv_feature = features_all[:, stream_start_index['vuv']:stream_start_index['vuv']+1]
for i in xrange(frame_number):
if vuv_feature[i, 0] < 0.5:
gen_features[i, 0] = inf_float
new_file_name = os.path.join(dir_name, file_id + file_ext_dict[feature_name])
dp.array_to_binary_file(gen_features, new_file_name)
print 'wrote to file %s' % new_file_name
pass
def acoustic_decomposition(self, in_file_list, dimension, out_dimension_dict, file_extension_dict, var_file_dict, do_MLPG=True, cfg=None):
print ('param_generation')
print ('acoustic_decomposition for %d files' % len(in_file_list) )
self.load_covariance(var_file_dict, out_dimension_dict)
stream_start_index = {}
dimension_index = 0
recorded_vuv = True
vuv_dimension = None
for feature_name in list(out_dimension_dict.keys()):
if feature_name != 'vuv':
stream_start_index[feature_name] = dimension_index
else:
vuv_dimension = dimension_index
dimension_index += out_dimension_dict[feature_name]
io_funcs = BinaryIOCollection()
mlpg_algo = MLParameterGeneration()
findex=0
flen=len(in_file_list)
for file_name in in_file_list:
findex=findex+1
dir_name = os.path.dirname(file_name)
file_id = os.path.splitext(os.path.basename(file_name))[0]
features, frame_number = io_funcs.load_binary_file_frame(file_name, dimension)
print('processing %4d of %4d: %s' % (findex,flen,file_name) )
for feature_name in self.gen_wav_features:
print(' feature: %s' % feature_name)
current_features = features[:, stream_start_index[feature_name]:stream_start_index[feature_name]+out_dimension_dict[feature_name]]
if FAST_MLPG:
### fast version wants variance per frame, not single global one:
var = self.var[feature_name]
var = numpy.transpose(numpy.tile(var,frame_number))
else:
var = self.var[feature_name]
# print var.shape[1]
if do_MLPG == False:
gen_features = current_features
else:
gen_features = mlpg_algo.generation(current_features, var, out_dimension_dict[feature_name]//3)
print(' feature dimensions: %d by %d' %(gen_features.shape[0], gen_features.shape[1]))
if feature_name in ['lf0', 'F0']:
if 'vuv' in stream_start_index:
vuv_feature = features[:, stream_start_index['vuv']:stream_start_index['vuv']+1]
for i in range(frame_number):
if new_vuv_feature[i] < 0.5:
gen_features[i, 0] = self.inf_float
new_file_name = os.path.join(dir_name, file_id + file_extension_dict[feature_name])
if self.enforce_silence:
silence_pattern = cfg.silence_pattern
label_align_dir = cfg.in_label_align_dir
in_f = open(label_align_dir+'/'+file_id+'.lab','r')
for line in in_f.readlines():
line = line.strip()
if len(line) < 1:
continue
temp_list = re.split('\s+', line)
start_time = int(int(temp_list[0])*(10**-4)/5)
end_time = int(int(temp_list[1])*(10**-4)/5)
full_label = temp_list[2]
label_binary_flag = self.check_silence_pattern(full_label, silence_pattern)
if label_binary_flag:
if feature_name in ['lf0', 'F0', 'mag']:
gen_features[start_time:end_time, :] = self.inf_float
else:
gen_features[start_time:end_time, :] = 0.0
io_funcs.array_to_binary_file(gen_features, new_file_name)
print(' wrote to file %s' % new_file_name)
def acoustic_decomposition(features,
frames,
file_id,
var_dict,
stream_start_index,
mlpg=True):
mlpg_algo = MLParameterGeneration()
#print(features.shape)
#print("frames:",frames)
#print(cfg.feats)
#print(stream_start_index)
for feature_name in stream_start_index:
#print("Not VUV!")
#print(feature_name)
end = stream_start_index[feature_name] + cfg.feats[feature_name]
#print(end)
current_features = features[:, stream_start_index[feature_name]:end]
#print("Avashna 0 - Before MLPG")
#print(current_features.shape)
#print("Merlin debug - Before MLPG")
#print(current_features[:10])
var = var_dict[feature_name]
#print("Debugging!", feature_name)
var = np.transpose(np.tile(var, frames))
#print("var", var)
if mlpg and feature_name not in ['vuv']:
#print("mlpg ")
gen_features = mlpg_algo.generation(current_features, var,
cfg.feats[feature_name] // 3)
filename = os.path.join(cfg.gen_path, file_id + "." + feature_name)
else:
#print("no mlpg")
gen_features = current_features
filename = os.path.join(cfg.gen_path, file_id + "." + feature_name)
#
#print("Avashna 1 - After MLPG")
#print(gen_features[:10]*-1.0e-9)
print(' feature dimensions: %d by %d' %
(gen_features.shape[0], gen_features.shape[1]))
if feature_name in ['lf0', 'F0']:
#print(gen_features[:100,])
#print("IN THIS LOOP")
if 'vuv' in stream_start_index:
#print("vuv present")
end = stream_start_index['vuv'] + cfg.feats['vuv']
vuv_feature = features[:, stream_start_index['vuv']:end]
#print("vuv",vuv_feature.shape)
#print(frames)
count = 0
for i in range(frames):
if vuv_feature[i, 0] < 0.5 or gen_features[i,
0] < np.log(20):
gen_features[i, 0] = -1.0e+10
count = count + 1
#print(count)
#print(gen_features[:100,])
#print("Avashna 2")
#print(gen_features)
save_binary(filename, gen_features)
print(' wrote to file %s' % filename)
def acoustic_decomposition(in_file_list,
dimension_all,
param_var_dir,
b_16k=True):
if b_16k: # 16k
param_dim_dict = {'mgc': 180, 'vuv': 1, 'lf0': 3, 'bap': 3}
else: # 24k
param_dim_dict = {'mgc': 180, 'vuv': 1, 'lf0': 3, 'bap': 9}
file_ext_dict = {'mgc': '.mgc', 'lf0': '.lf0', 'bap': '.bap'}
# load parameter variance
var_file_dict = {}
for f_name in param_dim_dict.keys():
var_file_dict[f_name] = os.path.join(
param_var_dir, f_name + '_' + str(param_dim_dict[f_name]))
var_dict = {}
for f_name in var_file_dict.keys():
var_value, _ = dap.load_binary_file_frame(var_file_dict[f_name], 1)
var_value = np.reshape(var_value, (param_dim_dict[f_name], 1))
var_dict[f_name] = var_value
# parameter start index
dimension_index = 0
stream_start_index = {}
for feature_name in param_dim_dict.keys():
stream_start_index[feature_name] = dimension_index
dimension_index += param_dim_dict[feature_name]
wave_feature_types = ['mgc', 'lf0', 'bap']
inf_float = -1.0e+10
mlpg = MLParameterGeneration()
# one cmp file per loop
for file_name in in_file_list:
dir_name = os.path.dirname(file_name)
file_id = os.path.splitext(os.path.basename(file_name))[0]
# load cmp data from file
features_all, frame_number = dap.load_binary_file_frame(
file_name, dimension_all)
# one type of features per loop
for feature_name in wave_feature_types:
curr_feature = features_all[:, stream_start_index[feature_name]:
stream_start_index[feature_name] +
param_dim_dict[feature_name]]
var = var_dict[feature_name]
var = np.transpose(np.tile(var, frame_number))
gen_features = mlpg.generation(curr_feature, var,
param_dim_dict[feature_name] / 3)
if feature_name in ['lf0', 'F0']:
if 'vuv' in stream_start_index:
vuv_feature = features_all[:, stream_start_index['vuv']:
stream_start_index['vuv'] + 1]
for i in xrange(frame_number):
if vuv_feature[i, 0] < 0.5:
gen_features[i, 0] = inf_float
new_file_name = os.path.join(dir_name,
file_id + file_ext_dict[feature_name])
dap.array_to_binary_file(gen_features, new_file_name)
print 'wrote to file %s' % new_file_name
pass