def cal_lf0(config):
base_path = config['base_path']
label_path = config['label_path']
name = config['name']
outfilepath = config['outfilepath']
var_path = config['var_path']
syllable_base_path = config['syllable_base_path']
syllable_var_path = config['syllable_var_path']
#----------Syllable level--------#
dur_list, names = PoGUtility.gen_dur_and_name_list(label_path, name)
# print dur_list
# print names
syl_mean = np.load('{}/mean.npy'.format(syllable_base_path))
# syl_mean, cccc = gen_mean_and_cov_of_dct_fake(names)
syl_cov = np.load('{}/cov.npy'.format(syllable_base_path))
print syl_cov
var = np.load('{}'.format(syllable_var_path))
vv = []
for i, v in enumerate(var):
vv.append(v[i])
syl_var = np.array(vv)
o = []
for data_dct, dur in zip(syl_mean, dur_list):
i_dct = PoGUtility.generate_inverse_DCT(data_dct, dur)
# print i_dct
o = o + i_dct
o = np.array(o)
o[o < 3] = np.nan
print o.shape
org = Utility.read_lf0_into_ascii(
'/work/w2/decha/Data/GPR_speccom_data/data_before_remove_silence/lf0/tsc/sd/j/{}.lf0'
.format(name))
org[org < 0] = np.nan
diff = len(org) - len(o)
plt.plot(np.arange(len(o)) + diff, o, label='syn')
plt.plot(range(len(org)), org, label='org')
plt.legend()
plt.savefig('./{}_dct_16_test.eps'.format(name))
sys.exit()
pass
def lf0_distortion(org_path, syn_path):
lf0_true_list = []
lf0_pred_list = []
for base in Utility.list_file(org_path):
if base.startswith('.'):
continue
# Load Original
original_file = os.path.join(org_path, base)
original_vector = numpy.loadtxt(
Utility.read_lf0_into_ascii(original_file))
# Load Synthesis
synthesis_file = os.path.join(syn_path, base)
synthesis_vector = numpy.loadtxt(
Utility.read_lf0_into_ascii(synthesis_file))
for lf0_original, lf0_synthesis in zip(original_vector,
synthesis_vector):
if lf0_original == Distortion.UNDEF_VALUE:
continue
if lf0_synthesis == Distortion.UNDEF_VALUE:
continue
lf0_true_list.append(lf0_original)
lf0_pred_list.append(lf0_synthesis)
rmse = numpy.sqrt(
sklearn.metrics.mean_squared_error(
lf0_true_list, lf0_pred_list)) * 1200 / numpy.log(2)
print('LF0 RMSE: {:f} in {} frames'.format(rmse, len(lf0_true_list)))
pass
def lf0_distortion_syn_is_gpr_format(org_path, syn_path):
lf0_true_list = []
lf0_pred_list = []
for base in Utility.list_file(org_path):
if base.startswith('.'):
continue
# if '12' in base: continue
# Load Original
original_file = os.path.join(org_path, base)
original_vector = numpy.loadtxt(
Utility.read_lf0_into_ascii(original_file))
# Load Synthesis
synthesis_file = '{}/{}.npy'.format(syn_path,
Utility.get_basefilename(base))
synthesis_vector = numpy.load(synthesis_file)
synthesis_vector = synthesis_vector.reshape(len(synthesis_vector))
# print synthesis_vector
#
synthesis_vector = np.nan_to_num(synthesis_vector)
synthesis_vector[np.where(synthesis_vector <= 0.0)] = UNDEF_VALUE
# print synthesis_vector
# sys.exit()
for lf0_original, lf0_synthesis in zip(original_vector,
synthesis_vector):
if lf0_original == UNDEF_VALUE:
continue
if lf0_synthesis == UNDEF_VALUE:
continue
lf0_true_list.append(lf0_original)
lf0_pred_list.append(lf0_synthesis)
rmse = numpy.sqrt(
sklearn.metrics.mean_squared_error(
lf0_true_list, lf0_pred_list)) * 1200 / numpy.log(2)
print('LF0 RMSE: {:f} in {} frames'.format(rmse, len(lf0_true_list)))
pass
def plot_syllable(lab_path, lf0_path, out_set_path, name, plot_set_out):
lines = Utility.read_file_line_by_line(lab_path)
lf0 = Utility.read_lf0_into_ascii(lf0_path)
print lf0
path = '{}/{}/'.format(out_set_path, name)
Utility.make_directory(path)
for idx, line in enumerate(lines):
line = Utility.trim(line)
spl = line.split(' ')
# print spl
start = float(spl[0]) / 50000
end = float(spl[1]) / 50000
syl = spl[2]
# print start, end, syl
if end > len(lf0):
end = len(lf0) - 1
cur_lf0 = lf0[start:end]
# print len(cur_lf0)
o = '{}_{}'.format((idx + 1), syl)
out_name = '{}/{}.lf0'.format(path, o)
print out_name
# Utility.write_to_file_line_by_line(out_name, cur_lf0)
Utility.make_directory('{}/{}/'.format(plot_set_out, name))
plot_out_file_path = '{}/{}/{}.eps'.format(plot_set_out, name, o)
plot(cur_lf0, plot_out_file_path)
print len(lf0)
pass
base_path = '{}/{}/'.format(frame_predicted_lf0_path, name)
label_path = '{}/{}.lab'.format(syl_duration_path, name)
var_path = '{}/inv_dimension_cov.npy'.format(
frame_predicted_lf0_path)
syllable_base_path = '{}/{}/'.format(syllable_predicted_dct_path,
name)
syllable_var_path = '{}/inv_dimension_cov.npy'.format(
syllable_predicted_dct_path)
vuv_path = '{}/{}/'.format(vuv_predicted_path, name)
original = Utility.read_lf0_into_ascii('{}/{}.lf0'.format(
original_path, name))
original = np.array(original)
original_vuv = np.copy(original)
original_vuv[original_vuv < 0] = -1
original_vuv[original_vuv >= 0] = +1
original[original < 0] = np.nan
koriyama_gen = np.load('{}/{}.npy'.format(koriyama_gen_path, name))
koriyama_gen[koriyama_gen < 0] = np.nan
stress = '{}/{}.npy'.format(stress_path, name)
stress = np.load(stress)
config = {
def lf0_distortion_syn_is_gpr_format(org_path, syn_path, stress_list,
mono_label):
lf0_true_list = []
lf0_pred_list = []
lf0_true_stress_list = []
lf0_pred_stress_list = []
for base in Utility.list_file(org_path):
if base.startswith('.'):
continue
b = Utility.get_basefilename(base)
stress = np.load('{}/{}.npy'.format(stress_list, b))
mono_file = Utility.read_file_line_by_line('{}/{}.lab'.format(
mono_label, b))
stress_index = np.array([])
for st, mono in zip(stress, mono_file):
spl = mono.split(' ')
start = int(spl[0]) / 50000
end = int(spl[1]) / 50000
if (st[0] == '1') & (st[1] == '{}'.format(tone)):
stress_index = np.append(stress_index,
np.arange(start, end),
axis=0)
# Load Original
original_file = os.path.join(org_path, base)
original_vector = numpy.loadtxt(
Utility.read_lf0_into_ascii(original_file))
# Load Synthesis
synthesis_file = '{}/{}.npy'.format(syn_path,
Utility.get_basefilename(base))
synthesis_vector = numpy.load(synthesis_file)
synthesis_vector = synthesis_vector.reshape(len(synthesis_vector))
# print synthesis_vector
synthesis_vector = np.nan_to_num(synthesis_vector)
synthesis_vector[np.where(synthesis_vector <= 0.0)] = UNDEF_VALUE
# print synthesis_vector
# sys.exit()
for idx, (lf0_original, lf0_synthesis) in enumerate(
zip(original_vector, synthesis_vector)):
if lf0_original == UNDEF_VALUE:
continue
if lf0_synthesis == UNDEF_VALUE:
continue
lf0_true_list.append(lf0_original)
lf0_pred_list.append(lf0_synthesis)
if idx in stress_index:
lf0_true_stress_list.append(lf0_original)
lf0_pred_stress_list.append(lf0_synthesis)
# rmse = numpy.sqrt(sklearn.metrics.mean_squared_error(lf0_true_list, lf0_pred_list)) * 1000 / numpy.log(2)
rmse = numpy.sqrt(
sklearn.metrics.mean_squared_error(
lf0_true_list, lf0_pred_list)) * 1200 / numpy.log(2)
print('All LF0 RMSE: {:f} in {} frames'.format(rmse, len(lf0_true_list)))
rmse = numpy.sqrt(
sklearn.metrics.mean_squared_error(
lf0_true_stress_list, lf0_pred_stress_list)) * 1200 / numpy.log(2)
print('Only stress LF0 RMSE: {:f} in {} frames'.format(
rmse, len(lf0_true_stress_list)))
pass
]
for path in paths:
for i in range(1, 51):
base = 'tscsdj{}'.format(Utility.fill_zero(i, 2))
lf0_single = np.load('{}/{}.npy'.format(path[0], base))
lf0_multi = np.load('{}/{}.npy'.format(path[1], base))
print base
single_vs_multi_rmse = lf0_distortion_syn_is_gpr_format(
lf0_single, lf0_multi)
original_lf0 = Utility.read_lf0_into_ascii('{}/{}.lf0'.format(
original, base))
single = lf0_distortion_syn_is_gpr_format(lf0_single, original_lf0)
multi = lf0_distortion_syn_is_gpr_format(lf0_multi, original_lf0)
print single_vs_multi_rmse, single, multi, 'Improve : ', (single -
multi)
plt.clf()
fig = plt.gcf()
fig.set_size_inches(15, 4)
original_lf0[original_lf0 < 0] = np.nan
lf0_single[lf0_single < 0] = np.nan
lf0_multi[lf0_multi < 0] = np.nan
from numpy.linalg import inv
if __name__ == '__main__':
unvoice = -1.00000000e+10
# syn = np.load('/work/w21/decha/Interspeech_2017/Result/01_Given_syllable_dct_Joint_probability/num_dct_cov_7/tscsdj01.npy')
method = '01_Given_syllable_model_combined_128'
filename = 'tscsdj01'
syn = np.load('/work/w21/decha/Interspeech_2017/Result/{}/num_dct_cov_7/{}.npy'.format(method, filename))
speech_param = np.load('/work/w16/decha/decha_w16/spec_com_work_space/Speech_synthesis/05a_GPR/testrun/out/tsc/a-i/speech_param/a-i/demo/seed-00/M-1024/B-1024/num_iters-5/lf0/param_mean/{}.npy'.format(filename))
org = Utility.read_lf0_into_ascii('/work/w2/decha/Data/GPR_speccom_data/data_before_remove_silence/lf0/tsc/sd/j/{}.lf0'.format(filename))
org = np.array(org)
idx = np.where(speech_param==unvoice)[0]
syn[idx] = np.nan
speech_param[idx] = np.nan
org[np.where(org==unvoice)[0]] = np.nan
x = range(len(syn))
fig = plt.gcf()
fig.set_size_inches(15, 4)
plt.plot(x , syn, label='Decha syn')
plt.plot(x , speech_param, label='Koriyama syn')
from tool_box.util.utility import Utility
from tool_box.distortion.distortion_utility import Distortion
from PoG_Utility.pog_utility import PoGUtility
import numpy as np
import matplotlib.pyplot as plt
from scipy.fftpack import dct, idct
if __name__ == '__main__':
org = '/work/w2/decha/Data/GPR_speccom_data/data_before_remove_silence/lf0/tsc/sd/j/tscsdj01.lf0'
org = Utility.read_lf0_into_ascii(org)
syn = '/work/w21/decha/Interspeech_2017/Result/From_03_with_mean_as_unvoice_lf0_format/num_dct_cov_7/tscsdj01.lf0'
syn = np.loadtxt(syn)
print syn[1000]
syn = '/work/w21/decha/Interspeech_2017/Result/From_01_lf0_format/num_dct_cov_7/tscsdj01.lf0'
syn = np.loadtxt(syn)
print syn[1000]
org[org==-1.00000000e+10] = np.nan
syn[syn==-1.00000000e+10] = np.nan
x = np.arange(len(org))
print x.shape, org.shape
def lf0_distortion_syn_is_gpr_format(org_path, data_dict, stress_list,
mono_label, tone, stress_type):
UNDEF_VALUE = -1.0e+10
lf0_true_list = []
lf0_pred_list = []
lf0_true_stress_list = []
lf0_pred_stress_list = []
for base in Utility.list_file(org_path):
if base.startswith('.'):
continue
b = Utility.get_basefilename(base)
stress = np.load('{}/{}.npy'.format(stress_list, b))
mono_file = Utility.read_file_line_by_line('{}/{}.lab'.format(
mono_label, b))
stress_index = np.array([])
# Load Synthesis
synthesis_vector = data_dict['initial'][1][b]
for st, mono in zip(stress, mono_file):
spl = mono.split(' ')
start = int(spl[0]) / 50000
end = int(spl[1]) / 50000
if not (st[0] == '1'):
st[0] = '0'
if (st[0] == str(stress_type)):
if str(st[2]) == '0':
pt = 'initial'
elif str(st[2]) == '1':
pt = 'vowel'
elif str(st[2]) == '2':
pt = 'final'
synthesis_vector[start:end] = data_dict[pt][int(
st[1])][b][start:end]
if '{}'.format(tone) == 'all':
stress_index = np.append(stress_index,
np.arange(start, end),
axis=0)
elif st[1] == '{}'.format(tone):
stress_index = np.append(stress_index,
np.arange(start, end),
axis=0)
# Load Original
original_file = os.path.join(org_path, base)
original_vector = np.loadtxt(
Utility.read_lf0_into_ascii(original_file))
# print synthesis_vector
synthesis_vector = np.nan_to_num(synthesis_vector)
synthesis_vector[np.where(synthesis_vector <= 0.0)] = UNDEF_VALUE
for idx, (lf0_original, lf0_synthesis) in enumerate(
zip(original_vector, synthesis_vector)):
if lf0_original == UNDEF_VALUE:
continue
if lf0_synthesis == UNDEF_VALUE:
continue
lf0_true_list.append(lf0_original)
lf0_pred_list.append(lf0_synthesis)
if idx in stress_index:
lf0_true_stress_list.append(lf0_original)
lf0_pred_stress_list.append(lf0_synthesis)
print 'Stress {}, Tone {}'.format(stress_type, tone)
rmse = np.sqrt(
sklearn.metrics.mean_squared_error(lf0_true_list,
lf0_pred_list)) * 1200 / np.log(2)
print('All LF0 RMSE: {:f} in {} frames'.format(rmse, len(lf0_true_list)))
rmse = np.sqrt(
sklearn.metrics.mean_squared_error(
lf0_true_stress_list, lf0_pred_stress_list)) * 1200 / np.log(2)
print('Only specific case LF0 RMSE: {:f} in {} frames'.format(
rmse, len(lf0_true_stress_list)))
return rmse
pass
def plot(lf0_path, label_path, filename, sett, intensity_path):
lf0 = Utility.read_lf0_into_ascii('{}/{}/{}.lf0'.format(
lf0_path, sett, filename))
lf0[lf0 < 0] = np.nan
intense_object = Utility.load_obj(intensity_path)
xtic_list = []
intensity_position = []
syl_ist = []
count = 0
# print '{}/{}/{}.lab'.format(label_path, sett, filename)
for line in Utility.read_file_line_by_line('{}/{}/{}.lab'.format(
label_path, sett, filename)):
spl = line.split(' ')
xtic_list.append(float(spl[1]) / 50000)
pos = (float(spl[1]) / 50000 + float(spl[0]) / 50000) / 2
text_pos = (float(spl[0]) / 50000) #+ float(spl[0])/50000) /2
intentsity = np.nan
if not (('-sil+' in spl[2]) | ('-pau+' in spl[2])):
count += 1
key = 'tscsd_manual_{}_{}'.format(filename[12:12 + 3], count)
intentsity = intense_object[key]
print intentsity
pattern = re.compile(
r"""
.+
/J:.+-(?P<consonant>.+)\+.+
/K:.+-(?P<vowel>.+)\+.+
/L:.+-(?P<finalconsonant>.+)\+.+""", re.VERBOSE)
match = re.match(pattern, spl[2])
if match:
syl = '{}-{}-{}'.format(match.group('consonant'),
match.group('vowel'),
match.group('finalconsonant'))
print syl
syl_ist.append((text_pos, syl))
intensity_position.append([pos, intentsity])
plt.plot(range(len(lf0)), lf0)
for x in xtic_list:
plt.plot([x, x], [plt.ylim()[0], plt.ylim()[1]], 'k--', lw=1)
for t in syl_ist:
plt.text(t[0],
plt.ylim()[1],
t[1],
fontsize=8,
rotation=45,
color='green')
ax2 = plt.twinx()
xx, yy = np.array([]), np.array([])
temp = 0
for p in intensity_position:
if len(yy) == 0:
temp = p[0]
else:
temp += p[0]
a = np.empty(p[0])
a.fill(p[1])
yy = np.append(yy, a)
print len(yy)
# print intensity_position
intensity_position = np.array(intensity_position)
xx = intensity_position[:, 0]
yy = intensity_position[:, 1]
ax2.plot(xx, yy, 'r-')
ylim = ax2.get_ylim()
ax2.set_ylim(ylim[0], 1.15)
plt.gcf().set_size_inches(14, 3)
plt.savefig('./test.eps')
pass
