def mod_power(self, cvmcep, rmcep, alpha=0.42, irlen=1024):
"""Power modification based on inpulse responce
Parameters
----------
cvmcep : array, shape (`T`, `dim`)
array of converted mel-cepstrum
rmcep : array, shape (`T`, `dim`)
array of reference mel-cepstrum
alpha : float, optional
All-path filter transfer function
Default set to 0.42
irlen : int, optional
Length for IIR filter
Default set to 1024
Return
------
modified_cvmcep : array, shape (`T`, `dim`)
array of power modified converted mel-cepstrum
"""
if rmcep.shape != cvmcep.shape:
raise ValueError("The shapes of the converted and \
reference mel-cepstrum are different: \
{} / {}".format(cvmcep.shape, rmcep.shape))
cv_e = pysptk.mc2e(cvmcep, alpha=alpha, irlen=irlen)
r_e = pysptk.mc2e(rmcep, alpha=alpha, irlen=irlen)
dpow = np.log(r_e / cv_e) / 2
modified_cvmcep = np.copy(cvmcep)
modified_cvmcep[:, 0] += dpow
return modified_cvmcep
def mod_p(m, r):
m1 = pysptk.mc2e(m, alpha=0.42, irlen=256)
r1 = pysptk.mc2e(r, alpha=0.42, irlen=256)
p = np.log(r1 / m1) / 2
t_m = np.copy(m)
t_m[:, 0] = t_m[:, 0] + p
return t_m
def mod_pow(cvmcep, mcep, alpha=MCEP_ALPHA, irlen=IRLEN):
cv_e = ps.mc2e(cvmcep, alpha=alpha, irlen=irlen)
r_e = ps.mc2e(mcep, alpha=alpha, irlen=irlen)
dpow = np.log(r_e / cv_e) / 2
mod_cvmcep = np.copy(cvmcep)
mod_cvmcep[:, 0] += dpow
return mod_cvmcep
def mod_power(cvmcep, rmcep, alpha, irlen=1024):
if rmcep.shape != cvmcep.shape:
raise ValueError("The shapes of the converted and \
reference mel-cepstrum are different: \
{} / {}".format(cvmcep.shape, rmcep.shape))
cv_e = pysptk.mc2e(cvmcep, alpha=alpha, irlen=irlen)
r_e = pysptk.mc2e(rmcep, alpha=alpha, irlen=irlen)
dpow = np.log(r_e / cv_e) / 2
modified_cvmcep = np.copy(cvmcep)
modified_cvmcep[:, 0] += dpow
return modified_cvmcep
def mod_power(cvmcep, rmcep, alpha=0.42, irlen=1024):
"""Power modification based on inpulse responce
Parameters
----------
cvmcep : array, shape (`T`, `dim`)
array of converted mel-cepstrum
rmcep : array, shape (`T`, `dim`)
array of reference mel-cepstrum
alpha : float, optional
All-path filter transfer function
Default set to 0.42
irlen : int, optional
Length for IIR filter
Default set to 1024
Return
------
modified_cvmcep : array, shape (`T`, `dim`)
array of power modified converted mel-cepstrum
"""
if rmcep.shape != cvmcep.shape:
raise ValueError("The shapes of the converted and \
reference mel-cepstrum are different: \
{} / {}".format(cvmcep.shape, rmcep.shape))
cv_e = pysptk.mc2e(cvmcep, alpha=alpha, irlen=irlen)
r_e = pysptk.mc2e(rmcep, alpha=alpha, irlen=irlen)
dpow = np.log(r_e / cv_e) / 2
modified_cvmcep = np.copy(cvmcep)
modified_cvmcep[:, 0] += dpow
return modified_cvmcep
def forward(self, audio, feat_kinds=['sp', 'mcc', 'f0', 'ap', 'en']):
"""Computes world features from a batch of waves
PARAMS
------
audio: Variable(torch.FloatTensor) with shape (T) in range [-1, 1]
RETURNS
-------
feat: torch.FloatTensor of shape ((SP+MCC+F0+AP+1+1), T)
Contains features in "feat_kinds": SP, MCC, F0, AP, SP_en, MCC_en
"""
device = audio.device
audio = audio.detach().cpu().numpy()
feat = dict()
for feat_kind in feat_kinds:
feat[feat_kind] = list()
for x in audio:
# Preprocess
x = x * MAX_WAV_VALUE
x = self.low_cut_filter(x, cutoff=self.cutoff_freq)
# Extract f0
f0, time_axis = pyworld.harvest(x,
self.fs,
f0_floor=self.minf0,
f0_ceil=self.maxf0,
frame_period=self.shiftms)
# Extract sp
sp = pyworld.cheaptrick(x,
f0,
time_axis,
self.fs,
fft_size=self.fft_size)
if 'sp' in feat_kinds:
feat['sp'].append(torch.from_numpy(sp).float().t())
# Extract ap
if 'ap' in feat_kinds:
ap = pyworld.d4c(x,
f0,
time_axis,
self.fs,
fft_size=self.fft_size)
feat['ap'].append(torch.from_numpy(ap).float().t())
# Extract mcc
if 'mcc' in feat_kinds:
mcc = pysptk.sp2mc(sp, self.mcc_dim, self.mcc_alpha)
feat['mcc'].append(torch.from_numpy(mcc).float().t())
# Extract energy
if 'en' in feat_kinds:
mcc = pysptk.sp2mc(sp, self.mcc_dim, self.mcc_alpha)
en = pysptk.mc2e(mcc, alpha=self.mcc_alpha, irlen=256)
# en = np.clip(en, 1e-10, None)
feat['en'].append(torch.from_numpy(en).float().view(-1))
# Fix f0
if 'f0' in feat_kinds:
f0[f0 < 0] = 0
feat['f0'].append(torch.from_numpy(f0).float().view(-1))
for key, val_list in feat.items():
feat[key] = torch.cat([val.unsqueeze(0) for val in val_list],
dim=0).to(device)
return feat
def test_mc2b():
x = windowed_dummy_data(1024)
mc = pysptk.mcep(x)
assert pysptk.mc2e(mc) > 0