Merge pull request #2 from Hiroshiba/harvest-realtime

Harvest realtime

1 files changed, 57 insertions, 7 deletions

diff --git a/become_yukarin/data_struct.py b/become_yukarin/data_struct.py
index 78c8cf3..4474331 100644
--- a/become_yukarin/data_struct.py
+++ b/become_yukarin/data_struct.py
@@ -1,8 +1,9 @@
-from typing import NamedTuple
+from typing import NamedTuple, Dict, List
 
+import numpy
 import pyworld
 
-import numpy
+_min_mc = -18.3
 
 
 class Wave(NamedTuple):
@@ -11,11 +12,21 @@ class Wave(NamedTuple):
 
 
 class AcousticFeature(NamedTuple):
-    f0: numpy.ndarray
-    spectrogram: numpy.ndarray
-    aperiodicity: numpy.ndarray
-    mfcc: numpy.ndarray
-    voiced: numpy.ndarray
+    f0: numpy.ndarray = numpy.nan
+    spectrogram: numpy.ndarray = numpy.nan
+    aperiodicity: numpy.ndarray = numpy.nan
+    mfcc: numpy.ndarray = numpy.nan
+    voiced: numpy.ndarray = numpy.nan
+
+    @staticmethod
+    def dtypes():
+        return dict(
+            f0=numpy.float32,
+            spectrogram=numpy.float32,
+            aperiodicity=numpy.float32,
+            mfcc=numpy.float32,
+            voiced=numpy.bool,
+        )
 
     def astype(self, dtype):
         return AcousticFeature(
@@ -51,6 +62,45 @@ class AcousticFeature(NamedTuple):
         assert self.voiced.dtype == numpy.bool
 
     @staticmethod
+    def silent(length: int, sizes: Dict[str, int], keys: List[str]):
+        d = {}
+        if 'f0' in keys:
+            d['f0'] = numpy.zeros((length, sizes['f0']), dtype=AcousticFeature.dtypes()['f0'])
+        if 'spectrogram' in keys:
+            d['spectrogram'] = numpy.zeros((length, sizes['spectrogram']),
+                                           dtype=AcousticFeature.dtypes()['spectrogram'])
+        if 'aperiodicity' in keys:
+            d['aperiodicity'] = numpy.zeros((length, sizes['aperiodicity']),
+                                            dtype=AcousticFeature.dtypes()['aperiodicity'])
+        if 'mfcc' in keys:
+            d['mfcc'] = numpy.hstack((
+                numpy.ones((length, 1), dtype=AcousticFeature.dtypes()['mfcc']) * _min_mc,
+                numpy.zeros((length, sizes['mfcc'] - 1), dtype=AcousticFeature.dtypes()['mfcc'])
+            ))
+        if 'voiced' in keys:
+            d['voiced'] = numpy.zeros((length, sizes['voiced']), dtype=AcousticFeature.dtypes()['voiced'])
+        feature = AcousticFeature(**d)
+        return feature
+
+    @staticmethod
+    def concatenate(fs: List['AcousticFeature'], keys: List[str]):
+        is_target = lambda a: not numpy.any(numpy.isnan(a))
+        return AcousticFeature(**{
+            key: numpy.concatenate([getattr(f, key) for f in fs]) if is_target(getattr(fs[0], key)) else numpy.nan
+            for key in keys
+        })
+
+    def pick(self, first: int, last: int):
+        is_target = lambda a: not numpy.any(numpy.isnan(a))
+        return AcousticFeature(
+            f0=self.f0[first:last] if is_target(self.f0) else numpy.nan,
+            spectrogram=self.spectrogram[first:last] if is_target(self.spectrogram) else numpy.nan,
+            aperiodicity=self.aperiodicity[first:last] if is_target(self.aperiodicity) else numpy.nan,
+            mfcc=self.mfcc[first:last] if is_target(self.mfcc) else numpy.nan,
+            voiced=self.voiced[first:last] if is_target(self.voiced) else numpy.nan,
+        )
+
+    @staticmethod
     def get_sizes(sampling_rate: int, order: int):
         fft_size = pyworld.get_cheaptrick_fft_size(fs=sampling_rate)
         return dict(