def __init__(self,
config: SRConfig,
model_path: Path,
gpu: int = None) -> None:
self.config = config
self.model_path = model_path
self.gpu = gpu
self.model = model = create_predictor_sr(config.model)
chainer.serializers.load_npz(str(model_path), model)
if self.gpu is not None:
model.to_gpu(self.gpu)
self._param = param = config.dataset.param
self._wave_process = WaveFileLoadProcess(
sample_rate=param.voice_param.sample_rate,
top_db=None,
)
self._low_high_spectrogram_process = LowHighSpectrogramFeatureProcess(
frame_period=param.acoustic_feature_param.frame_period,
order=param.acoustic_feature_param.order,
alpha=param.acoustic_feature_param.alpha,
f0_estimating_method=param.acoustic_feature_param.
f0_estimating_method,
)
self._low_high_spectrogram_load_process = LowHighSpectrogramFeatureLoadProcess(
validate=True, )
def generate_file(path):
out = Path(arguments.output_directory, path.stem + '.npy')
if out.exists() and not arguments.enable_overwrite:
return
# load wave and padding
wave_file_load_process = WaveFileLoadProcess(
sample_rate=arguments.sample_rate,
top_db=arguments.top_db,
pad_second=arguments.pad_second,
)
wave = wave_file_load_process(path, test=True)
# make acoustic feature
acoustic_feature_process = AcousticFeatureProcess(
frame_period=arguments.frame_period,
order=arguments.order,
alpha=arguments.alpha,
f0_estimating_method=arguments.f0_estimating_method,
)
feature = acoustic_feature_process(wave, test=True).astype_only_float(numpy.float32)
high_spectrogram = feature.spectrogram
fftlen = pyworld.get_cheaptrick_fft_size(arguments.sample_rate)
low_spectrogram = pysptk.mc2sp(
feature.mfcc,
alpha=arguments.alpha,
fftlen=fftlen,
)
# save
numpy.save(out.absolute(), {
'low': low_spectrogram,
'high': high_spectrogram,
})
def process(p: Path, super_resolution: SuperResolution):
param = config.dataset.param
wave_process = WaveFileLoadProcess(
sample_rate=param.voice_param.sample_rate,
top_db=None,
)
acoustic_feature_process = AcousticFeatureProcess(
frame_period=param.acoustic_feature_param.frame_period,
order=param.acoustic_feature_param.order,
alpha=param.acoustic_feature_param.alpha,
f0_estimating_method=param.acoustic_feature_param.f0_estimating_method,
)
try:
if p.suffix in ['.npy', '.npz']:
p = glob.glob(str(input_wave_directory / p.stem) + '.*')[0]
p = Path(p)
input = acoustic_feature_process(wave_process(str(p)))
wave = super_resolution(input.spectrogram,
acoustic_feature=input,
sampling_rate=param.voice_param.sample_rate)
librosa.output.write_wav(str(output / p.stem) + '.wav',
wave.wave,
wave.sampling_rate,
norm=True)
except:
import traceback
print('error!', str(p))
traceback.format_exc()
def __init__(self,
config: Config,
model_path: Path,
gpu: int = None) -> None:
self.config = config
self.model_path = model_path
self.gpu = gpu
self.model = model = create_predictor(config.model)
chainer.serializers.load_npz(str(model_path), model)
if self.gpu is not None:
model.to_gpu(self.gpu)
self._param = param = config.dataset.param
self._wave_process = WaveFileLoadProcess(
sample_rate=param.voice_param.sample_rate,
top_db=None,
)
self._feature_process = AcousticFeatureProcess(
frame_period=param.acoustic_feature_param.frame_period,
order=param.acoustic_feature_param.order,
alpha=param.acoustic_feature_param.alpha,
f0_estimating_method=param.acoustic_feature_param.
f0_estimating_method,
)
self._acoustic_feature_load_process = acoustic_feature_load_process = AcousticFeatureLoadProcess(
)
input_mean = acoustic_feature_load_process(
config.dataset.input_mean_path, test=True)
input_var = acoustic_feature_load_process(
config.dataset.input_var_path, test=True)
target_mean = acoustic_feature_load_process(
config.dataset.target_mean_path, test=True)
target_var = acoustic_feature_load_process(
config.dataset.target_var_path, test=True)
self._feature_normalize = AcousticFeatureNormalizeProcess(
mean=input_mean,
var=input_var,
)
self._feature_denormalize = AcousticFeatureDenormalizeProcess(
mean=target_mean,
var=target_var,
)
feature_sizes = AcousticFeature.get_sizes(
sampling_rate=param.voice_param.sample_rate,
order=param.acoustic_feature_param.order,
)
self._encode_feature = EncodeFeatureProcess(config.dataset.features)
self._decode_feature = DecodeFeatureProcess(config.dataset.features,
feature_sizes)
def generate_feature(path1, path2):
out1 = Path(arguments.output1_directory, path1.stem + '.npy')
out2 = Path(arguments.output2_directory, path2.stem + '.npy')
if out1.exists() and out2.exists() and not arguments.enable_overwrite:
return
# load wave and padding
wave_file_load_process = WaveFileLoadProcess(
sample_rate=arguments.sample_rate,
top_db=arguments.top_db,
pad_second=arguments.pad_second,
)
wave1 = wave_file_load_process(path1, test=True)
wave2 = wave_file_load_process(path2, test=True)
# make acoustic feature
acoustic_feature_process1 = AcousticFeatureProcess(
frame_period=arguments.frame_period,
order=arguments.order,
alpha=arguments.alpha,
f0_estimating_method=arguments.f0_estimating_method,
f0_floor=arguments.f0_floor1,
f0_ceil=arguments.f0_ceil1,
)
acoustic_feature_process2 = AcousticFeatureProcess(
frame_period=arguments.frame_period,
order=arguments.order,
alpha=arguments.alpha,
f0_estimating_method=arguments.f0_estimating_method,
f0_floor=arguments.f0_floor2,
f0_ceil=arguments.f0_ceil2,
)
f1 = acoustic_feature_process1(wave1,
test=True).astype_only_float(numpy.float32)
f2 = acoustic_feature_process2(wave2,
test=True).astype_only_float(numpy.float32)
# pre convert
if pre_convert:
f1_ref = pre_converter1.convert_to_feature(f1)
else:
f1_ref = f1
# alignment
if not arguments.disable_alignment:
aligner = MFCCAligner(f1_ref.mfcc, f2.mfcc)
f0_1, f0_2 = aligner.align(f1.f0, f2.f0)
spectrogram_1, spectrogram_2 = aligner.align(f1.spectrogram,
f2.spectrogram)
aperiodicity_1, aperiodicity_2 = aligner.align(f1.aperiodicity,
f2.aperiodicity)
mfcc_1, mfcc_2 = aligner.align(f1.mfcc, f2.mfcc)
voiced_1, voiced_2 = aligner.align(f1.voiced, f2.voiced)
f1 = AcousticFeature(
f0=f0_1,
spectrogram=spectrogram_1,
aperiodicity=aperiodicity_1,
mfcc=mfcc_1,
voiced=voiced_1,
)
f2 = AcousticFeature(
f0=f0_2,
spectrogram=spectrogram_2,
aperiodicity=aperiodicity_2,
mfcc=mfcc_2,
voiced=voiced_2,
)
f1.validate()
f2.validate()
# save
acoustic_feature_save_process = AcousticFeatureSaveProcess(
validate=True, ignore=arguments.ignore_feature)
acoustic_feature_save_process({'path': out1, 'feature': f1})
print('saved!', out1)
acoustic_feature_save_process({'path': out2, 'feature': f2})
print('saved!', out2)