align without eliminate

2 files changed, 19 insertions, 3 deletions

diff --git a/become_yukarin/dataset/dataset.py b/become_yukarin/dataset/dataset.py
index b049ed8..0ca35df 100644
--- a/become_yukarin/dataset/dataset.py
+++ b/become_yukarin/dataset/dataset.py
@@ -73,7 +73,8 @@ class WaveFileLoadProcess(BaseDataProcess):
 
     def __call__(self, data: str, test):
         wave = librosa.core.load(data, sr=self._sample_rate, dtype=self._dtype)[0]
-        wave = librosa.effects.remix(wave, intervals=librosa.effects.split(wave, top_db=self._top_db))
+        if self._top_db is not None:
+            wave = librosa.effects.remix(wave, intervals=librosa.effects.split(wave, top_db=self._top_db))
         return Wave(wave, self._sample_rate)
 
 
diff --git a/become_yukarin/dataset/utility.py b/become_yukarin/dataset/utility.py
index b2f5480..c28b1df 100644
--- a/become_yukarin/dataset/utility.py
+++ b/become_yukarin/dataset/utility.py
@@ -42,5 +42,20 @@ class DTWAligner(object):
 
 
 class MFCCAligner(DTWAligner):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, dist=nnmnkwii.metrics.melcd, **kwargs)
+    def __init__(self, x, y, *args, **kwargs):
+        x = self._calc_aligner_feature(x)
+        y = self._calc_aligner_feature(y)
+        super().__init__(x, y, *args, dist=nnmnkwii.metrics.melcd, **kwargs)
+
+    @classmethod
+    def _calc_delta(cls, x):
+        x = numpy.zeros_like(x, x.dtype)
+        x[:-1] = x[:-1] - x[1:]
+        x[-1] = 0
+        return x
+
+    @classmethod
+    def _calc_aligner_feature(cls, x):
+        d = cls._calc_delta(x)
+        feature = numpy.concatenate((x, d), axis=1)[:, 1:]
+        return feature