リアルタイム変換が可能

3 files changed, 25 insertions, 8 deletions

diff --git a/become_yukarin/dataset/dataset.py b/become_yukarin/dataset/dataset.py
index 178844a..7e3acbf 100644
--- a/become_yukarin/dataset/dataset.py
+++ b/become_yukarin/dataset/dataset.py
@@ -87,8 +87,16 @@ class WaveFileLoadProcess(BaseDataProcess):
 
 
 class AcousticFeatureProcess(BaseDataProcess):
-    def __init__(self, frame_period, order, alpha, f0_estimating_method, f0_floor=71, f0_ceil=800,
-                 dtype=numpy.float32) -> None:
+    def __init__(
+            self,
+            frame_period,
+            order,
+            alpha,
+            f0_estimating_method,
+            f0_floor=71,
+            f0_ceil=800,
+            dtype=numpy.float32,
+    ) -> None:
         self._frame_period = frame_period
         self._order = order
         self._alpha = alpha
@@ -110,7 +118,8 @@ class AcousticFeatureProcess(BaseDataProcess):
                 f0_ceil=self._f0_ceil,
             )
         else:
-            _f0, t = pyworld.harvest(
+            from world4py.np import apis
+            _f0, t = apis.harvest(
                 x,
                 fs,
                 frame_period=self._frame_period,
diff --git a/become_yukarin/vocoder.py b/become_yukarin/vocoder.py
index f1a9f03..0674e0f 100644
--- a/become_yukarin/vocoder.py
+++ b/become_yukarin/vocoder.py
@@ -65,7 +65,7 @@ class RealtimeVocoder(Vocoder):
             number_of_pointers,  # number of pointers
             self._synthesizer,
         )
-        self._before_buffer = None  # for holding memory
+        self._before_buffer = []  # for holding memory
 
     def decode(
             self,
@@ -93,7 +93,9 @@ class RealtimeVocoder(Vocoder):
                 sampling_rate=self.out_sampling_rate,
             )
 
-        self._before_buffer = (f0_buffer, sp_buffer, ap_buffer)  # for holding memory
+        self._before_buffer.append((f0_buffer, sp_buffer, ap_buffer))  # for holding memory
+        if len(self._before_buffer) > 16:
+            self._before_buffer.pop(0)
         return out_wave
 
     def warm_up(self, time_length: float):
diff --git a/become_yukarin/voice_changer.py b/become_yukarin/voice_changer.py
index 05f5a96..5e0eac0 100644
--- a/become_yukarin/voice_changer.py
+++ b/become_yukarin/voice_changer.py
@@ -83,7 +83,7 @@ class VoiceChangerStream(object):
     def remove_wave(self, end_time: float):
         self._data_stream = list(filter(lambda s: s.end_time > end_time, self._data_stream))
 
-    def convert(self, start_time: float, time_length: float):
+    def convert_to_feature(self, start_time: float, time_length: float):
         end_time = start_time + time_length
         buffer_list = []
         stream = filter(lambda s: not (end_time < s.start_time or s.end_time < start_time), self._data_stream)
@@ -121,12 +121,18 @@ class VoiceChangerStream(object):
             buffer_list.append(pad)
 
         buffer = numpy.concatenate(buffer_list)
-        print('buffer', len(buffer), flush=True)
         in_wave = Wave(wave=buffer, sampling_rate=self.sampling_rate)
         in_feature = self.vocoder.encode(in_wave)
         out_feature = self.voice_changer.convert_from_acoustic_feature(in_feature)
         return out_feature
 
+    def convert(self, start_time: float, time_length: float):
+        feature = self.convert_to_feature(start_time=start_time, time_length=time_length)
+        out_wave = self.vocoder.decode(
+            acoustic_feature=feature,
+        )
+        return out_wave
+
     def convert_with_extra_time(self, start_time: float, time_length: float, extra_time: float):
         """
         :param extra_time: 音声変換時に余分に使うデータの時間長。ゼロパディングを防ぐ。
@@ -136,7 +142,7 @@ class VoiceChangerStream(object):
         start_time -= extra_time
         time_length += extra_time * 2
 
-        extra_feature = self.convert(start_time=start_time, time_length=time_length)
+        extra_feature = self.convert_to_feature(start_time=start_time, time_length=time_length)
 
         pad = int(extra_time / (frame_period / 1000))
         feature = AcousticFeature(