diff options
| -rw-r--r-- | become_yukarin/dataset/dataset.py | 1 | ||||
| -rw-r--r-- | become_yukarin/param.py | 2 | ||||
| -rw-r--r-- | become_yukarin/voice_changer.py | 13 |
3 files changed, 11 insertions, 5 deletions
diff --git a/become_yukarin/dataset/dataset.py b/become_yukarin/dataset/dataset.py index 4fded11..6328a1c 100644 --- a/become_yukarin/dataset/dataset.py +++ b/become_yukarin/dataset/dataset.py @@ -192,7 +192,6 @@ class AcousticFeatureDenormalizeProcess(BaseDataProcess): def __call__(self, data: AcousticFeature, test): f0 = data.f0 * numpy.sqrt(self._var.f0) + self._mean.f0 - print(data.voiced.dtype) f0[~data.voiced] = 0 return AcousticFeature( f0=f0, diff --git a/become_yukarin/param.py b/become_yukarin/param.py index db49c0a..529378c 100644 --- a/become_yukarin/param.py +++ b/become_yukarin/param.py @@ -3,7 +3,7 @@ from typing import NamedTuple class VoiceParam(NamedTuple): sample_rate: int = 24000 - top_db: float = 20 + top_db: float = 40 class AcousticFeatureParam(NamedTuple): diff --git a/become_yukarin/voice_changer.py b/become_yukarin/voice_changer.py index c160a5d..935b95c 100644 --- a/become_yukarin/voice_changer.py +++ b/become_yukarin/voice_changer.py @@ -77,6 +77,13 @@ class VoiceChanger(object): out = self.model(inputs).data[0] out = self._decode_feature(out, test=True) + out = AcousticFeature( + f0=out.f0, + spectrogram=numpy.nan, + aperiodicity=input_feature.aperiodicity, + mfcc=out.mfcc, + voiced=input_feature.voiced, + ) out = self._feature_denormalize(out, test=True) fftlen = pyworld.get_cheaptrick_fft_size(input_wave.sampling_rate) @@ -89,12 +96,12 @@ class VoiceChanger(object): out = AcousticFeature( f0=out.f0, spectrogram=spectrogram, - aperiodicity=input_feature.aperiodicity, + aperiodicity=out.aperiodicity, mfcc=out.mfcc, - voiced=input_feature.voiced, + voiced=out.voiced, ).astype(numpy.float64) out = pyworld.synthesize( - f0=out.f0, + f0=out.f0.ravel(), spectrogram=out.spectrogram, aperiodicity=out.aperiodicity, fs=out_sampling_rate, |