def process(p_in: Path, acoustic_converter: AcousticConverter):
try:
if p_in.suffix in ['.npy', '.npz']:
p_in = Path(glob.glob(str(dataset_wave_dir / p_in.stem) + '.*')[0])
# input wave
w_in = acoustic_converter.load_wave(p_in)
w_in.wave *= input_scale
# input feature
f_in = acoustic_converter.extract_acoustic_feature(w_in)
f_in_effective, effective = acoustic_converter.separate_effective(
wave=w_in, feature=f_in, threshold=threshold)
# convert
f_out = acoustic_converter.convert_loop(f_in_effective)
f_out = acoustic_converter.combine_silent(effective=effective,
feature=f_out)
f_out = acoustic_converter.decode_spectrogram(f_out)
# save
sampling_rate = acoustic_converter.out_sampling_rate
frame_period = acoustic_converter.config.dataset.acoustic_param.frame_period
wave = f_out.decode(sampling_rate=sampling_rate,
frame_period=frame_period)
librosa.output.write_wav(y=wave.wave,
path=str(output_dir / (p_in.stem + '.wav')),
sr=wave.sampling_rate)
except:
import traceback
traceback.print_exc()
def _convert(p_in: Path, acoustic_converter: AcousticConverter,
super_resolution: SuperResolution):
w_in = acoustic_converter.load_wave(p_in)
f_in = acoustic_converter.extract_acoustic_feature(w_in)
f_in_effective, effective = acoustic_converter.separate_effective(
wave=w_in, feature=f_in)
f_low = acoustic_converter.convert(f_in_effective)
f_low = acoustic_converter.combine_silent(effective=effective,
feature=f_low)
f_low = acoustic_converter.decode_spectrogram(f_low)
s_high = super_resolution.convert(f_low.sp.astype(numpy.float32))
f_low_sr = BYAcousticFeature(
f0=f_low.f0,
spectrogram=f_low.sp,
aperiodicity=f_low.ap,
mfcc=f_low.mc,
voiced=f_low.voiced,
)
rate = acoustic_converter.out_sampling_rate
wave = super_resolution(s_high,
acoustic_feature=f_low_sr,
sampling_rate=rate)
return wave
def process(p_in: Path, acoustic_converter: AcousticConverter,
super_resolution: SuperResolution):
try:
if p_in.suffix in ['.npy', '.npz']:
p_in = Path(
glob.glob(str(dataset_input_wave_dir / p_in.stem) + '.*')[0])
w_in = acoustic_converter.load_wave(p_in)
f_in = acoustic_converter.extract_acoustic_feature(w_in)
f_in_effective, effective = acoustic_converter.separate_effective(
wave=w_in, feature=f_in)
f_low = acoustic_converter.convert(f_in_effective)
f_low = acoustic_converter.combine_silent(effective=effective,
feature=f_low)
if filter_size is not None:
f_low.f0 = AcousticConverter.filter_f0(f_low.f0,
filter_size=filter_size)
f_low = acoustic_converter.decode_spectrogram(f_low)
s_high = super_resolution.convert(f_low.sp.astype(numpy.float32))
# target
paths = glob.glob(str(dataset_target_wave_dir / p_in.stem) + '.*')
has_true = len(paths) > 0
if has_true:
p_true = Path(paths[0])
w_true = acoustic_converter.load_wave(p_true)
f_true = acoustic_converter.extract_acoustic_feature(w_true)
# save figure
fig = plt.figure(figsize=[36, 22])
plt.subplot(4, 1, 1)
plt.imshow(numpy.log(f_in.sp).T, aspect='auto', origin='reverse')
plt.plot(f_in.f0, 'w')
plt.colorbar()
plt.subplot(4, 1, 2)
plt.imshow(numpy.log(f_low.sp).T, aspect='auto', origin='reverse')
plt.plot(f_low.f0, 'w')
plt.colorbar()
plt.subplot(4, 1, 3)
plt.imshow(numpy.log(s_high).T, aspect='auto', origin='reverse')
plt.colorbar()
if has_true:
plt.subplot(4, 1, 4)
plt.imshow(numpy.log(f_true.sp).T, aspect='auto', origin='reverse')
plt.plot(f_true.f0, 'w')
plt.colorbar()
fig.savefig(output / (p_in.stem + '.png'))
# save wave
f_low_sr = BYAcousticFeature(
f0=f_low.f0,
spectrogram=f_low.sp,
aperiodicity=f_low.ap,
mfcc=f_low.mc,
voiced=f_low.voiced,
)
rate = acoustic_converter.out_sampling_rate
wave = super_resolution(s_high,
acoustic_feature=f_low_sr,
sampling_rate=rate)
librosa.output.write_wav(y=wave.wave,
path=str(output / (p_in.stem + '.wav')),
sr=rate)
except:
pass