無音を無視するように変更

4 files changed, 17 insertions, 8 deletions

diff --git a/become_yukarin/config/config.py b/become_yukarin/config/config.py
index f1f24cf..ed128c4 100644
--- a/become_yukarin/config/config.py
+++ b/become_yukarin/config/config.py
@@ -145,3 +145,6 @@ def backward_compatible(d: Dict):
         d['model']['generator_extensive_layers'] = 8
         d['model']['discriminator_base_channels'] = 32
         d['model']['discriminator_extensive_layers'] = 5
+
+    if 'weak_discriminator' not in d['model']:
+        d['model']['weak_discriminator'] = False
diff --git a/become_yukarin/voice_changer.py b/become_yukarin/voice_changer.py
index bed155f..7f7bbe4 100644
--- a/become_yukarin/voice_changer.py
+++ b/become_yukarin/voice_changer.py
@@ -4,6 +4,7 @@ from typing import NamedTuple
 
 import numpy
 
+from become_yukarin.param import Param
 from .acoustic_converter import AcousticConverter
 from .data_struct import AcousticFeature
 from .data_struct import Wave
@@ -78,10 +79,12 @@ class VoiceChangerStream(object):
             self,
             sampling_rate: int,
             frame_period: float,
+            order: int,
             in_dtype=numpy.float32,
     ):
         self.sampling_rate = sampling_rate
         self.frame_period = frame_period
+        self.order = order
         self.in_dtype = in_dtype
 
         self.voice_changer: VoiceChanger = None
@@ -189,8 +192,7 @@ class VoiceChangerStream(object):
         return in_feature
 
     def convert(self, start_time: float, time_length: float, extra_time: float):
-        order = self.voice_changer.acoustic_converter.config.dataset.param.acoustic_feature_param.order
-        sizes = AcousticFeature.get_sizes(sampling_rate=self.sampling_rate, order=order)
+        sizes = AcousticFeature.get_sizes(sampling_rate=self.sampling_rate, order=self.order)
         keys = ['f0', 'aperiodicity', 'mfcc', 'voiced']
         in_feature = self.fetch(
             start_time=start_time,
@@ -209,8 +211,7 @@ class VoiceChangerStream(object):
         return out_feature
 
     def post_convert(self, start_time: float, time_length: float):
-        order = self.voice_changer.acoustic_converter.config.dataset.param.acoustic_feature_param.order
-        sizes = AcousticFeature.get_sizes(sampling_rate=self.sampling_rate, order=order)
+        sizes = AcousticFeature.get_sizes(sampling_rate=self.sampling_rate, order=self.order)
         keys = ['f0', 'aperiodicity', 'spectrogram', 'voiced']
         out_feature = self.fetch(
             start_time=start_time,
diff --git a/scripts/realtime_voice_changer.py b/scripts/realtime_voice_changer.py
index e96ce4e..bda64dd 100644
--- a/scripts/realtime_voice_changer.py
+++ b/scripts/realtime_voice_changer.py
@@ -1,11 +1,9 @@
+import librosa
 import world4py
 
 world4py._WORLD_LIBRARY_PATH = 'x64_world.dll'
 
-from functools import partial
 from pathlib import Path
-import signal
-import time
 from typing import NamedTuple
 from multiprocessing import Queue
 from multiprocessing import Process
@@ -34,6 +32,7 @@ class AudioConfig(NamedTuple):
     convert_chunk: int
     vocoder_buffer_size: int
     out_norm: float
+    silent_threshold: float
 
 
 def encode_worker(
@@ -123,7 +122,10 @@ def decode_worker(
         wave_fragment = numpy.concatenate([wave_fragment, wave])
         if len(wave_fragment) >= audio_config.audio_chunk:
             wave, wave_fragment = wave_fragment[:audio_config.audio_chunk], wave_fragment[audio_config.audio_chunk:]
-            queue_output.put(wave)
+
+            power = librosa.core.power_to_db(numpy.abs(librosa.stft(wave)) ** 2).mean()
+            if power >= audio_config.silent_threshold:
+                queue_output.put(wave)
 
 
 def main():
@@ -154,11 +156,13 @@ def main():
         convert_chunk=config.dataset.param.voice_param.sample_rate,
         vocoder_buffer_size=config.dataset.param.voice_param.sample_rate // 16,
         out_norm=2.5,
+        silent_threshold=-99.0,
     )
 
     voice_changer_stream = VoiceChangerStream(
         sampling_rate=audio_config.rate,
         frame_period=config.dataset.param.acoustic_feature_param.frame_period,
+        order=config.dataset.param.acoustic_feature_param.order,
         in_dtype=numpy.float32,
     )
 
diff --git a/tests/test_voice_changer.py b/tests/test_voice_changer.py
index 66ea003..9772bbe 100644
--- a/tests/test_voice_changer.py
+++ b/tests/test_voice_changer.py
@@ -67,6 +67,7 @@ voice_changer = VoiceChanger(
 voice_changer_stream = VoiceChangerStream(
     sampling_rate=audio_config.rate,
     frame_period=acoustic_converter._param.acoustic_feature_param.frame_period,
+    order=acoustic_converter._param.acoustic_feature_param.order,
     in_dtype=numpy.float32,
 )