def _decode_feature(self, data):
sizes = AcousticFeature.get_sizes(
sampling_rate=self._param.sampling_rate,
order=self._param.order,
)
return decode_feature(data,
targets=self.config.dataset.out_features,
sizes=sizes)
def combine_silent(self, effective: numpy.ndarray, feature: AcousticFeature):
sizes = AcousticFeature.get_sizes(
sampling_rate=self._param.sampling_rate,
order=self._param.order,
)
silent_feature = AcousticFeature.silent(len(effective), sizes=sizes, keys=('mc', 'ap', 'f0', 'voiced'))
silent_feature.indexing_set(effective, feature)
return silent_feature
def pad(self, width: int):
sizes = AcousticFeature.get_sizes(
sampling_rate=self.wave_sampling_rate, order=self.order)
return AcousticFeatureWrapper.silent_wrapper(
width,
sizes=sizes,
keys=self._keys,
frame_period=self.frame_period,
sampling_rate=self.wave_sampling_rate,
wave_dtype=numpy.float32,
).astype_only_float_wrapper(numpy.float32)
def post_convert(self, start_time: float, time_length: float):
sizes = AcousticFeature.get_sizes(sampling_rate=self.sampling_rate, order=self.order)
keys = ['f0', 'ap', 'sp', 'voiced']
out_feature = self.fetch(
start_time=start_time,
time_length=time_length,
data_stream=self._out_feature_stream,
rate=1000 / self.frame_period,
pad_function=lambda length: AcousticFeature.silent(length, sizes=sizes, keys=keys),
pick_function=lambda segment, first, last: segment.feature.pick(first, last, keys=keys),
concat_function=lambda buffers: AcousticFeature.concatenate(buffers, keys=keys),
)
out_wave = self.vocoder.decode(
acoustic_feature=out_feature,
)
w = out_wave.wave
w[numpy.isnan(w)] = 0
out_wave = Wave(wave=w, sampling_rate=out_wave.sampling_rate)
return out_wave
def convert(self, start_time: float, time_length: float, extra_time: float):
sizes = AcousticFeature.get_sizes(sampling_rate=self.sampling_rate, order=self.order)
keys = ['f0', 'ap', 'mc', 'voiced']
def _pad_function(length):
return AcousticFeatureWrapper.silent_wrapper(
length,
sizes=sizes,
keys=keys,
frame_period=self.frame_period,
sampling_rate=self.sampling_rate,
wave_dtype=self.in_dtype,
).astype_only_float_wrapper(self.in_dtype)
def _pick_function(segment: FeatureWrapperSegment, first, last):
return segment.feature.pick_wrapper(
first,
last,
keys=keys,
frame_period=self.frame_period,
)
in_feature = self.fetch(
start_time=start_time,
time_length=time_length,
extra_time=extra_time,
data_stream=self._in_feature_stream,
rate=1000 / self.frame_period,
pad_function=_pad_function,
pick_function=_pick_function,
concat_function=lambda buffers: AcousticFeatureWrapper.concatenate_wrapper(buffers, keys=keys),
)
out_feature = self.voice_changer.convert_from_acoustic_feature(in_feature)
pad = round(extra_time * 1000 / self.frame_period)
out_feature = out_feature.pick(pad, -pad, keys=['f0', 'ap', 'sp', 'voiced'])
return out_feature
def pad(self, width: int):
sizes = AcousticFeature.get_sizes(
sampling_rate=self.wave_sampling_rate, order=self.order)
return AcousticFeature.silent(width, sizes=sizes, keys=self._keys)