def gen_W(number_of_frame, dur_list, num_coeff):
w = np.zeros((num_coeff * len(dur_list), number_of_frame))
offset_x = 0
offset_y = 0
for idx, d in enumerate(dur_list):
if idx == (len(dur_list) - 1):
d = number_of_frame - offset_x
local_w = PoGUtility.generate_W_for_DCT(d, num_coeff)
# print offset_x, offset_y, local_w.shape
for i in range(num_coeff):
w[offset_y + i][offset_x:offset_x + d] = local_w[i]
offset_x = offset_x + d
offset_y = offset_y + num_coeff
return w
pass
head_count, tail_count = find_head_and_tail(phone['raw_lf0'])
raw = np.array(phone['raw_lf0'])[head_count : len(phone['raw_lf0'])-tail_count]
# print raw
if len(raw) < 10:
continue
else:
if len(raw[raw<0]) != 0:
# print raw
raw = Utility.inteporate_lf0(raw)
# print raw
phone['clean_raw'] = raw
phone_clean_dict[phone_id] = phone
w = PoGUtility.generate_W_for_DCT(len(raw), coeff)
data_dct = np.dot(w, raw)
# print data_dct
phone['dct'] = data_dct
# sys.exit()
print len(phone_clean_dict)
Utility.save_obj(phone_clean_dict, new_phone_dict)
pass
def gen_W(number_of_frame,
dur_list,
num_coeff,
stress_list,
vuv,
tone,
stress_type,
p_in_s_file,
use_consonant=False):
# print stress_list
w = np.zeros((num_coeff * len(dur_list), number_of_frame))
offset_x = 0
offset_y = 0
for idx, (d, p_in_s,
st) in enumerate(zip(dur_list, p_in_s_file, stress_list)):
if idx == (len(dur_list) - 1):
d = number_of_frame - offset_x
if use_consonant:
if len(p_in_s) == 0:
consonant = 0
else:
consonant = int(float(p_in_s[0]) / float(50000.0))
cur_vuv = vuv[offset_x + consonant:offset_x + d]
head, tail = PoGUtility.find_head_and_tail(cur_vuv)
if (((consonant + tail + head) > d) | (not (st[0] == '1')) |
(not (st[1] == '{}'.format(tone)))):
local_w = np.zeros((num_coeff, d))
else:
# print 'panda'
voice_frame = d - consonant - tail - head
local_w = PoGUtility.generate_W_for_DCT(
voice_frame, num_coeff)
if head != 0:
local_w = np.concatenate((np.zeros(
(num_coeff, head), dtype='float'), local_w),
axis=1)
if tail != 0:
local_w = np.concatenate(
(local_w, np.zeros(
(num_coeff, tail), dtype='float')),
axis=1)
if consonant != 0:
local_w = np.concatenate((np.zeros(
(num_coeff, consonant), dtype='float'), local_w),
axis=1)
else:
cur_vuv = vuv[offset_x:offset_x + d]
head, tail = PoGUtility.find_head_and_tail(cur_vuv)
if not (st[0] == '1'):
st[0] = '0'
if (((tail + head) > d) | (not (st[0] == str(stress_type))) |
(not (st[1] == '{}'.format(tone)))):
local_w = np.zeros((num_coeff, d))
else:
# print 'panda'
voice_frame = d - tail - head
local_w = PoGUtility.generate_W_for_DCT(
voice_frame, num_coeff)
if head != 0:
local_w = np.concatenate((np.zeros(
(num_coeff, head), dtype='float'), local_w),
axis=1)
if tail != 0:
local_w = np.concatenate(
(local_w, np.zeros(
(num_coeff, tail), dtype='float')),
axis=1)
for i in range(num_coeff):
w[offset_y + i][offset_x:offset_x + d] = local_w[i]
offset_x = offset_x + d
offset_y = offset_y + num_coeff
return w
errors = dict()
errors_list = []
errors_tuple = []
true = np.array([])
dct_regen = np.array([])
for coeff in [3, 4, 7]:
for name in d:
data = d[name]
w = PoGUtility.generate_W_for_DCT(len(data), coeff)
data_dct = PoGUtility.generate_DCT(data, coeff)
data_dct = np.dot(w, data)
i_dct = PoGUtility.generate_inverse_DCT(data_dct, len(data))
true = np.concatenate((true, data))
dct_regen = np.concatenate((dct_regen, i_dct))
rmse = np.sqrt(sklearn.metrics.mean_squared_error(
data, i_dct)) * 1200 / np.log(2)
# print rmse
errors[name] = rmse
errors_list.append(rmse)
from scipy.fftpack import dct, idct
import math
if __name__ == '__main__':
syllable_dict = Utility.load_obj(
'/work/w2/decha/Data/GPR_speccom_data/Interspeech2017/syllable_dictionary_data_with_delta_deltadelta.pkl'
)
for syl in syllable_dict:
print syl
lf0 = syllable_dict[syl][0]
y = 0.0
for n, f in enumerate(lf0):
y = y + f * math.cos(math.pi * 0 * (2.0 * n + 1) /
(2.0 * len(lf0)))
W = PoGUtility.generate_W_for_DCT(len(lf0), len(lf0))
lf0_dct = dct(lf0, norm='ortho')
print lf0_dct
sys.exit()
pass
