def generate_align_indexes(pair_path: Tuple[Path, Path]):
path1, path2 = pair_path
if path1.stem != path2.stem:
print('warning: the file names are different', path1, path2)
out = Path(arguments.output, path1.stem + '.npy')
if out.exists() and not arguments.enable_overwrite:
return
# original
wave = Wave.load(path=path1, sampling_rate=sconf1.wav_fs)
wave = wave.pad(pre_second=arguments.pad_second1,
post_second=arguments.pad_second1)
x = low_cut_filter(wave.wave, wave.sampling_rate, cutoff=70)
feat1.analyze(x)
mcep = feat1.mcep(dim=sconf1.mcep_dim, alpha=sconf1.mcep_alpha)
if arguments.threshold_db1 is not None:
indexes = wave.get_effective_frame(
threshold_db=arguments.threshold_db1,
fft_length=sconf1.wav_fftl,
frame_period=sconf1.wav_shiftms,
)
mcep = mcep[indexes]
cvmcep_wopow = mcepgmm.convert(static_delta(mcep[:, 1:]),
cvtype=pconf.GMM_mcep_cvtype)
mcep1 = numpy.c_[mcep[:, 0], cvmcep_wopow]
# target
wave = Wave.load(path=path2, sampling_rate=sconf2.wav_fs)
wave = wave.pad(pre_second=arguments.pad_second2,
post_second=arguments.pad_second2)
x = low_cut_filter(wave.wave, wave.sampling_rate, cutoff=70)
feat2.analyze(x)
mcep2 = feat2.mcep(dim=sconf2.mcep_dim, alpha=sconf2.mcep_alpha)
if arguments.threshold_db2 is not None:
indexes = wave.get_effective_frame(
threshold_db=arguments.threshold_db2,
fft_length=sconf2.wav_fftl,
frame_period=sconf2.wav_shiftms,
)
mcep2 = mcep2[indexes]
# align
feature1 = AcousticFeature(mc=mcep1)
feature2 = AcousticFeature(mc=mcep2)
align_indexes = AlignIndexes.extract(feature1,
feature2,
dtype=arguments.dtype)
align_indexes.save(path=out,
validate=True,
ignores=arguments.ignore_feature)
def convert(self, in_feature: AcousticFeature):
im, iv = self.input_statistics.mean, self.input_statistics.var
tm, tv = self.target_statistics.mean, self.target_statistics.var
f0 = numpy.copy(in_feature.f0)
f0[f0.nonzero()] = numpy.exp((tv / iv) * (numpy.log(f0[f0.nonzero()]) - im) + tm)
return AcousticFeature(f0=f0)
def make_mask(feature: AcousticFeature):
return AcousticFeature(
f0=feature.voiced,
sp=numpy.ones_like(feature.sp, dtype=numpy.bool),
ap=numpy.ones_like(feature.ap, dtype=numpy.bool),
coded_ap=numpy.ones_like(feature.coded_ap, dtype=numpy.bool),
mc=numpy.ones_like(feature.mc, dtype=numpy.bool),
voiced=numpy.ones_like(feature.voiced, dtype=numpy.bool),
).astype(numpy.float32)
def decode_feature(data: numpy.ndarray, targets: List[str], sizes: Dict[str, int]):
data = data.T
lasts = numpy.cumsum([sizes[t] for t in targets]).tolist()
assert data.shape[1] == lasts[-1]
return AcousticFeature(**{
t: data[:, bef:aft]
for t, bef, aft in zip(targets, [0] + lasts[:-1], lasts)
})