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 calc_score(path: Path):
scores = []
wave = Wave.load(path=path, sampling_rate=sconf.wav_fs)
wave = wave.pad(pre_second=arguments.pad_second,
post_second=arguments.pad_second)
hop = sconf.wav_fs * sconf.wav_shiftms // 1000
length = int(math.ceil(len(wave.wave) / hop + 0.0001))
# for sprocket
x = low_cut_filter(wave.wave, wave.sampling_rate, cutoff=70)
feat.analyze(x)
npow = feat.npow()
effective1: numpy.ndarray = (npow > sconf.power_threshold)
assert len(effective1) == length, str(path)
# for yukarin
for th in arguments.candidate_threshold:
effective2 = wave.get_effective_frame(
threshold_db=th,
fft_length=sconf.wav_fftl,
frame_period=sconf.wav_shiftms,
)
scores.append([
(effective1 == effective2).sum(),
(effective1 == effective2)[effective1].sum(),
(effective1 == effective2)[~effective1].sum(),
length,
])
return scores
def get_segments(self, values: Iterable[float],
time_lengths: Iterable[float]):
return AcousticFeatureWrapper(
wave=Wave(
wave=numpy.concatenate([
numpy.ones(round(time_length * self.wave_sampling_rate),
dtype=numpy.float32) * value
for value, time_length in zip(values, time_lengths)
]),
sampling_rate=self.wave_sampling_rate,
),
f0=numpy.concatenate([
numpy.ones((round(time_length * self.sampling_rate), 1),
dtype=numpy.float32) * value
for value, time_length in zip(values, time_lengths)
]),
)
def _encode(self, w: numpy.ndarray):
wave = Wave(wave=w, sampling_rate=self.input_rate)
feature_wrapper = self.realtime_vocoder.encode(wave)
return feature_wrapper