def get_feature(wav_path, preprocessing=False, getsize=False):
fs, x = wavfile.read(wav_path)
x = x.astype(np.float64)
if audio_world_config.use_harvest:
f0, timeaxis = pyworld.harvest(
x,
fs,
frame_period=audio_world_config.frame_period,
f0_floor=audio_world_config.f0_floor,
f0_ceil=audio_world_config.f0_ceil)
else:
f0, timeaxis = pyworld.dio(
x,
fs,
frame_period=audio_world_config.frame_period,
f0_floor=audio_world_config.f0_floor,
f0_ceil=audio_world_config.f0_ceil)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
bap = pyworld.code_aperiodicity(aperiodicity, fs)
alpha = pysptk.util.mcepalpha(fs)
mgc = pysptk.sp2mc(spectrogram,
order=audio_world_config.mgc_dim,
alpha=alpha)
f0 = f0[:, None]
lf0 = f0.copy()
nonzero_indices = np.nonzero(f0)
lf0[nonzero_indices] = np.log(f0[nonzero_indices])
if audio_world_config.use_harvest:
# https://github.com/mmorise/World/issues/35#issuecomment-306521887
vuv = (aperiodicity[:, 0] < 0.5).astype(np.float32)[:, None]
else:
vuv = (lf0 != 0).astype(np.float32)
lf0 = P.interp1d(lf0, kind=audio_world_config.f0_interpolation_kind)
# Parameter trajectory smoothing
if audio_world_config.mod_spec_smoothing:
hop_length = int(fs * (audio_world_config.frame_period * 0.001))
modfs = fs / hop_length
mgc = P.modspec_smoothing(
mgc, modfs, cutoff=audio_world_config.mod_spec_smoothing_cutoff)
mgc = P.delta_features(mgc, audio_world_config.windows)
lf0 = P.delta_features(lf0, audio_world_config.windows)
bap = P.delta_features(bap, audio_world_config.windows)
features = np.hstack((mgc, lf0, vuv, bap))
if preprocessing:
out_path = wav_path.replace(".wav", "").replace("wav", "world")
np.save(out_path, features)
elif getsize:
feature, mgc.shape[0], lf0.shape[0], bap.shape[0]
else:
return features
def collect_features(self, wav_path, label_path):
fs, x = wavfile.read(wav_path)
x = x.astype(np.float64)
if hp_acoustic.use_harvest:
f0, timeaxis = pyworld.harvest(
x, fs, frame_period=hp_acoustic.frame_period,
f0_floor=hp_acoustic.f0_floor, f0_ceil=hp_acoustic.f0_ceil)
else:
f0, timeaxis = pyworld.dio(
x, fs, frame_period=hp_acoustic.frame_period,
f0_floor=hp_acoustic.f0_floor, f0_ceil=hp_acoustic.f0_ceil)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
bap = pyworld.code_aperiodicity(aperiodicity, fs)
if self.alpha is None:
self.alpha = pysptk.util.mcepalpha(fs)
mgc = pysptk.sp2mc(spectrogram, order=hp_acoustic.order, alpha=self.alpha)
f0 = f0[:, None]
lf0 = f0.copy()
nonzero_indices = np.nonzero(f0)
lf0[nonzero_indices] = np.log(f0[nonzero_indices])
if hp_acoustic.use_harvest:
# https://github.com/mmorise/World/issues/35#issuecomment-306521887
vuv = (aperiodicity[:, 0] < 0.5).astype(np.float32)[:, None]
else:
vuv = (lf0 != 0).astype(np.float32)
lf0 = P.interp1d(lf0, kind=hp_acoustic.f0_interpolation_kind)
# Parameter trajectory smoothing
if hp_acoustic.mod_spec_smoothing:
hop_length = int(fs * (hp_acoustic.frame_period * 0.001))
modfs = fs / hop_length
mgc = P.modspec_smoothing(
mgc, modfs, cutoff=hp_acoustic.mod_spec_smoothing_cutoff)
mgc = P.delta_features(mgc, hp_acoustic.windows)
lf0 = P.delta_features(lf0, hp_acoustic.windows)
bap = P.delta_features(bap, hp_acoustic.windows)
features = np.hstack((mgc, lf0, vuv, bap))
# Cut silence frames by HTS alignment
labels = hts.load(label_path)
features = features[:labels.num_frames()]
indices = labels.silence_frame_indices()
features = np.delete(features, indices, axis=0)
return features.astype(np.float32)
def generate_changed_voice(model, input_path):
fs, x = wavfile.read(input_path)
x = x.astype(np.float64)
if len(x.shape) > 1:
x = x.mean(axis=1)
f0, timeaxis = pyworld.dio(x, fs, frame_period=hp.frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
alpha = pysptk.util.mcepalpha(fs)
mc = pysptk.sp2mc(spectrogram, order=hp.order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
mc = P.modspec_smoothing(mc, FS / HOP_LENGHT, cutoff=50)
mc = P.delta_features(mc, hp.windows).astype(np.float32)
gen_data = model.predict(mc)
gen_data = np.hstack([c0.reshape((-1, 1)), gen_data])
fftlen = pyworld.get_cheaptrick_fft_size(fs)
spectrogram = pysptk.mc2sp(
gen_data.astype(np.float64), alpha=alpha, fftlen=fftlen)
waveform = pyworld.synthesize(
f0, spectrogram, aperiodicity, fs, hp.frame_period)
return waveform
def test_delta_features():
T = 5
static_dim = 2
x = np.random.rand(T, static_dim)
for windows in _get_windows_set():
y = delta_features(x, windows)
assert y.shape == (T, static_dim * len(windows))
def function(self, feature, raw, key):
if self.frame_period is None:
self.frame_period = raw.frame_period
elif self.frame_period != raw.frame_period:
raise ValueError(f'frame_period of "{key}" is {raw.frame_period!r}'
f' but others are {self.frame_period!r}')
return delta_features(feature, DELTA_WINDOWS)
def test_vc_from_path(model, path, data_mean, data_std, diffvc=True):
model.eval()
fs, x = wavfile.read(path)
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=hp.frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
alpha = pysptk.util.mcepalpha(fs)
mc = pysptk.sp2mc(spectrogram, order=hp.order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
static_dim = mc.shape[-1]
mc = P.modspec_smoothing(mc, fs / hop_length, cutoff=50)
mc = P.delta_features(mc, hp.windows).astype(np.float32)
T = mc.shape[0]
inputs = mc[:, :static_dim].copy()
# Normalization
mc_scaled = P.scale(mc, data_mean, data_std)
# Apply model
mc_scaled = Variable(torch.from_numpy(mc_scaled))
R = unit_variance_mlpg_matrix(hp.windows, T)
R = torch.from_numpy(R)
y_hat, y_hat_static = model(mc_scaled, R)
mc_static_pred = y_hat_static.data.cpu().numpy().reshape(-1, static_dim)
# Denormalize
mc_static_pred = P.inv_scale(mc_static_pred, data_mean[:static_dim],
data_std[:static_dim])
outputs = mc_static_pred.copy()
if diffvc:
mc_static_pred = mc_static_pred - mc[:, :static_dim]
mc = np.hstack((c0[:, None], mc_static_pred))
if diffvc:
mc[:, 0] = 0 # remove power coefficients
engine = Synthesizer(MLSADF(order=hp.order, alpha=alpha),
hopsize=hop_length)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
else:
fftlen = pyworld.get_cheaptrick_fft_size(fs)
spectrogram = pysptk.mc2sp(mc.astype(np.float64),
alpha=alpha,
fftlen=fftlen)
waveform = pyworld.synthesize(f0, spectrogram, aperiodicity, fs,
hp.frame_period)
return waveform, inputs, outputs
def test_delta_dataset():
base = {
'fp5': feature.get_analyzer(dataset.CLB_WAV),
'fp5_2': feature.get_analyzer(dataset.CLB_WAV2),
'fp3': feature.get_analyzer(dataset.CLB_WAV, frame_period=3)
}
mcep_dataset = MelCepstrumDataset(base)
delta_dataset = DeltaFeatureDataset(mcep_dataset)
assert (delta_dataset['fp5'] == delta_features(mcep_dataset['fp5'],
DELTA_WINDOWS)).all()
assert (delta_dataset['fp5_2'] == delta_features(mcep_dataset['fp5_2'],
DELTA_WINDOWS)).all()
with pytest.raises(ValueError) as e:
delta_dataset['fp3']
assert 'frame_period of "fp3" is 3 but others are 5' == str(e.value)
def collect_features(self, wav_path, label_path):
fs, x = wavfile.read(wav_path)
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x,
fs,
frame_period=hp_acoustic.frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
bap = pyworld.code_aperiodicity(aperiodicity, fs)
if self.alpha is None:
self.alpha = pysptk.util.mcepalpha(fs)
mgc = pysptk.sp2mc(spectrogram,
order=hp_acoustic.order,
alpha=self.alpha)
f0 = f0[:, None]
lf0 = f0.copy()
nonzero_indices = np.nonzero(f0)
lf0[nonzero_indices] = np.log(f0[nonzero_indices])
vuv = (lf0 != 0).astype(np.float32)
lf0 = P.interp1d(lf0, kind=hp_acoustic.f0_interpolation_kind)
# 50hz parameter trajectory smoothing
hop_length = int(fs * (hp_acoustic.frame_period * 0.001))
modfs = fs / hop_length
mgc = P.modspec_smoothing(mgc, modfs, cutoff=50)
mgc = P.delta_features(mgc, hp_acoustic.windows)
lf0 = P.delta_features(lf0, hp_acoustic.windows)
bap = P.delta_features(bap, hp_acoustic.windows)
features = np.hstack((mgc, lf0, vuv, bap))
# Cut silence frames by HTS alignment
labels = hts.load(label_path)
features = features[:labels.num_frames()]
indices = labels.silence_frame_indices()
features = np.delete(features, indices, axis=0)
return features.astype(np.float32)
def convert(self, feature, raw, **kwargs):
if self.frame_period != raw.frame_period:
raise ValueError(f'frame_period is expected to'
f' {self.frame_period!s}'
f' but {raw.frame_period!s}')
dim = feature.shape[-1]
result = super().convert(delta_features(feature, DELTA_WINDOWS),
**kwargs)
if result.shape[-1] > dim:
result = result[:, :dim]
return result
def recompute_delta_features(Y, Y_data_mean, Y_data_std,
windows,
stream_sizes=[180, 3, 1, 3],
has_dynamic_features=[True, True, False, True]):
start_indices = np.hstack(([0], np.cumsum(stream_sizes)[:-1]))
end_indices = np.cumsum(stream_sizes)
static_stream_sizes = get_static_stream_sizes(
stream_sizes, has_dynamic_features, len(windows))
for start_idx, end_idx, static_size, has_dynamic in zip(
start_indices, end_indices, static_stream_sizes, has_dynamic_features):
if has_dynamic:
y_static = Y[:, start_idx:start_idx + static_size]
Y[:, start_idx:end_idx] = P.delta_features(y_static, windows)
return Y
def collect_features(self, wav_path):
fs, x = wavfile.read(wav_path)
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=hp.frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
spectrogram = P.trim_zeros_frames(spectrogram)
if self.alpha is None:
self.alpha = pysptk.util.mcepalpha(fs)
mgc = pysptk.sp2mc(spectrogram, order=hp.order, alpha=self.alpha)
# Drop 0-th coefficient
mgc = mgc[:, 1:]
# 50Hz cut-off MS smoothing
hop_length = int(fs * (hp.frame_period * 0.001))
modfs = fs / hop_length
mgc = P.modspec_smoothing(mgc, modfs, cutoff=50)
# Add delta
mgc = P.delta_features(mgc, hp.windows)
return mgc.astype(np.float32)
def collect_features(self, wav_path):
fs, x = wavfile.read(wav_path)
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=hp.frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
spectrogram = P.trim_zeros_frames(spectrogram)
if self.alpha is None:
self.alpha = pysptk.util.mcepalpha(fs)
mgc = pysptk.sp2mc(spectrogram, order=hp.order, alpha=self.alpha)
# Drop 0-th coefficient
mgc = mgc[:, 1:]
# 50Hz cut-off MS smoothing
hop_length = int(fs * (hp.frame_period * 0.001))
modfs = fs / hop_length
mgc = P.modspec_smoothing(mgc, modfs, cutoff=50)
# Add delta
mgc = P.delta_features(mgc, hp.windows)
return mgc.astype(np.float32)
def recompute_delta_features(Y,
Y_data_mean,
Y_data_std,
windows,
stream_sizes=[180, 3, 1, 3],
has_dynamic_features=[True, True, False, True]):
start_indices = np.hstack(([0], np.cumsum(stream_sizes)[:-1]))
end_indices = np.cumsum(stream_sizes)
static_stream_sizes = get_static_stream_sizes(stream_sizes,
has_dynamic_features,
len(windows))
for start_idx, end_idx, static_size, has_dynamic in zip(
start_indices, end_indices, static_stream_sizes,
has_dynamic_features):
if has_dynamic:
y_static = Y[:, start_idx:start_idx + static_size]
Y[:, start_idx:end_idx] = P.delta_features(y_static, windows)
return Y
def test_one_utt(src_path, tgt_path, disable_mlpg=False, diffvc=True):
# GMM-based parameter generation is provided by the library in `baseline` module
if disable_mlpg:
# Force disable MLPG
paramgen = MLPG(gmm, windows=[(0, 0, np.array([1.0]))], diff=diffvc)
else:
paramgen = MLPG(gmm, windows=windows, diff=diffvc)
fs, x = wavfile.read(src_path)
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
pdb.set_trace()
mc = pysptk.sp2mc(spectrogram, order=order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
if use_delta:
mc = delta_features(mc, windows)
mc = paramgen.transform(mc)
if disable_mlpg and mc.shape[-1] != static_dim:
mc = mc[:, :static_dim]
assert mc.shape[-1] == static_dim
mc = np.hstack((c0[:, None], mc))
if diffvc:
mc[:, 0] = 0 # remove power coefficients
engine = Synthesizer(MLSADF(order=order, alpha=alpha),
hopsize=hop_length)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
else:
spectrogram = pysptk.mc2sp(mc.astype(np.float64),
alpha=alpha,
fftlen=fftlen)
waveform = pyworld.synthesize(f0, spectrogram, aperiodicity, fs,
frame_period)
return waveform
def test_one_utt(path_src, path_tgt, disable_mlpg=False, diffvc=True):
if disable_mlpg:
paramgen = MLPG(gmm, windows=[(0, 0, np.array([1.0]))], diff=diffvc)
else:
paramgen = MLPG(gmm, windows=windows, diff=diffvc)
x, fs_ = sf.read(path_src)
x = x.astype(np.float64)
f0, time_axis = pyworld.dio(x, fs_, frame_period=frame_period)
f0 = pyworld.stonemask(x, f0, time_axis, fs_)
spectrogram = pyworld.cheaptrick(x, f0, time_axis, fs_)
aperiodicity = pyworld.d4c(x, f0, time_axis, fs_)
mc = pysptk.sp2mc(spectrogram, order=order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
if use_delta:
mc = delta_features(mc, windows)
mc = paramgen.transform(mc)
if disable_mlpg and mc.shape[-1] != static_dim:
mc = mc[:, :static_dim]
assert mc.shape[-1] == static_dim
mc = np.hstack((c0[:, None], mc))
if diffvc:
mc[:, 0] = 0
engine = Synthesizer(MLSADF(order=order, alpha=alpha),
hopsize=hop_length)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
else:
spectrogram = pysptk.mc2sp(mc.astype(np.float64),
alpha=alpha,
fftlen=fftlen)
waveform = pyworld.synthesize(f0, spectrogram, aperiodicity, fs_,
frame_period)
return waveform
def test_vc_from_path(model, x, fs, data_mean, data_std, diffvc=True):
model.eval()
hop_length = int(fs * (hp.frame_period * 0.001))
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=hp.frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
alpha = pysptk.util.mcepalpha(fs)
mc = pysptk.sp2mc(spectrogram, order=hp.order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
static_dim = mc.shape[-1]
mc = P.modspec_smoothing(mc, fs / hop_length, cutoff=50)
mc = P.delta_features(mc, hp.windows).astype(np.float32)
T = mc.shape[0]
inputs = mc[:, :static_dim].copy()
# Normalization
mc_scaled = P.scale(mc, data_mean, data_std)
mc_scaled = Variable(torch.from_numpy(mc_scaled))
lengths = [len(mc_scaled)]
# Add batch axis
mc_scaled = mc_scaled.view(1, -1, mc_scaled.size(-1))
# For MLPG
R = unit_variance_mlpg_matrix(hp.windows, T)
R = torch.from_numpy(R)
# Apply model
if model.include_parameter_generation():
# Case: models include parameter generation in itself
# Mulistream features cannot be used in this case
y_hat, y_hat_static = model(mc_scaled, R, lengths=lengths)
else:
# Case: generic models (can be sequence model)
assert hp.has_dynamic_features is not None
y_hat = model(mc_scaled, lengths=lengths)
y_hat_static = multi_stream_mlpg(y_hat, R, hp.stream_sizes,
hp.has_dynamic_features)
mc_static_pred = y_hat_static.data.cpu().numpy().reshape(-1, static_dim)
# Denormalize
mc_static_pred = P.inv_scale(mc_static_pred, data_mean[:static_dim],
data_std[:static_dim])
outputs = mc_static_pred.copy()
if diffvc:
mc_static_pred = mc_static_pred - mc[:, :static_dim]
mc = np.hstack((c0[:, None], mc_static_pred))
if diffvc:
mc[:, 0] = 0 # remove power coefficients
engine = Synthesizer(MLSADF(order=hp.order, alpha=alpha),
hopsize=hop_length)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
else:
fftlen = pyworld.get_cheaptrick_fft_size(fs)
spectrogram = pysptk.mc2sp(mc.astype(np.float64),
alpha=alpha,
fftlen=fftlen)
waveform = pyworld.synthesize(f0, spectrogram, aperiodicity, fs,
hp.frame_period)
return waveform, inputs, outputs
# Parameter generation
paramgen = MLPG(gmm, windows=windows, diff=True)
# Waveform generation for test set
for idx, path in enumerate(source.test_paths):
fs, x = wavfile.read(path)
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=frame_period)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
# aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
mc = pysptk.sp2mc(spectrogram, order=order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
mc = delta_features(mc, windows)
since = time.time()
mc = paramgen.transform(mc)
print("{}, Elapsed time in conversion: {}s".format(idx,
time.time() - since))
assert mc.shape[-1] == static_dim
mc = np.hstack((c0[:, None], mc))
mc[:, 0] = 0
engine = Synthesizer(MLSADF(order=order, alpha=alpha), hopsize=80)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
if not exists('resultsVC'):
os.makedirs('resultsVC')
wavfile.write(
"resultsVC/{}_{}.wav".format(splitext(basename(path))[0], 'mlpg'), fs,
F=5 # Fold No
'''
for i in np.arange(0,460):
E_t,M_t,TT=Get_Wav_EMA_PerFile(EMAfiles[i],Wavfiles[i],i)
#W_t=W_t#[np.newaxis,:,:,:]
E_t=E_t#[np.newaxis,:,:]
M_t=M_t#[np.newaxis,:,:]
Youttrain.append(E_t)
X_trainseq.append(M_t)
TT_Train.append(TT)
'''
for i in np.arange(0,460):
if (((i+F)%10)==0):# Test
E_t,M_t,TT=Get_Wav_EMA_PerFile(EMAfiles[i],Wavfiles[i],i)
E_t= delta_features(E_t, windows)
M_t= delta_features(M_t, windows)
#W_t=W_t[np.newaxis,:,:,np.newaxis]
E_t=E_t[np.newaxis,:,:]
M_t=M_t[np.newaxis,:,:]
Youttest.append(E_t)
X_testseq.append(M_t)
TT_Test.append(TT)
#print('Test '+str(i))
elif (((i+F+1)%10)==0):# Validation
E_t,M_t,TT=Get_Wav_EMA_PerFile(EMAfiles[i],Wavfiles[i],i)
E_t= delta_features(E_t, windows)
M_t= delta_features(M_t, windows)
#W_t=W_t#[np.newaxis,:,:,:]
E_t=E_t#[np.newaxis,:,:]
M_t=M_t #forward_process(M_t[:,:n_mcep])
def test_vc_from_path(model, x, fs, data_mean, data_std, diffvc=True):
model.eval()
hop_length = int(fs * (hp.frame_period * 0.001))
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=hp.frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
alpha = pysptk.util.mcepalpha(fs)
mc = pysptk.sp2mc(spectrogram, order=hp.order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
static_dim = mc.shape[-1]
mc = P.modspec_smoothing(mc, fs / hop_length, cutoff=50)
mc = P.delta_features(mc, hp.windows).astype(np.float32)
T = mc.shape[0]
inputs = mc[:, :static_dim].copy()
# Normalization
mc_scaled = P.scale(mc, data_mean, data_std)
mc_scaled = Variable(torch.from_numpy(mc_scaled))
lengths = [len(mc_scaled)]
# Add batch axis
mc_scaled = mc_scaled.view(1, -1, mc_scaled.size(-1))
# For MLPG
R = unit_variance_mlpg_matrix(hp.windows, T)
R = torch.from_numpy(R)
# Apply model
if model.include_parameter_generation():
# Case: models include parameter generation in itself
# Mulistream features cannot be used in this case
y_hat, y_hat_static = model(mc_scaled, R, lengths=lengths)
else:
# Case: generic models (can be sequence model)
assert hp.has_dynamic_features is not None
y_hat = model(mc_scaled, lengths=lengths)
y_hat_static = multi_stream_mlpg(
y_hat, R, hp.stream_sizes, hp.has_dynamic_features)
mc_static_pred = y_hat_static.data.cpu().numpy().reshape(-1, static_dim)
# Denormalize
mc_static_pred = P.inv_scale(
mc_static_pred, data_mean[:static_dim], data_std[:static_dim])
outputs = mc_static_pred.copy()
if diffvc:
mc_static_pred = mc_static_pred - mc[:, :static_dim]
mc = np.hstack((c0[:, None], mc_static_pred))
if diffvc:
mc[:, 0] = 0 # remove power coefficients
engine = Synthesizer(MLSADF(order=hp.order, alpha=alpha),
hopsize=hop_length)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
else:
fftlen = pyworld.get_cheaptrick_fft_size(fs)
spectrogram = pysptk.mc2sp(
mc.astype(np.float64), alpha=alpha, fftlen=fftlen)
waveform = pyworld.synthesize(
f0, spectrogram, aperiodicity, fs, hp.frame_period)
return waveform, inputs, outputs
