Al Lot

audio_toolset/audio_io.py

@@ -76,9 +76,9 @@ class NormalizeMelband(object):
 
 
 class AudioToMel(object):
-    def __init__(self, amplitude_to_db=False, power_to_db=False, **kwargs):
+    def __init__(self, amplitude_to_db=False, power_to_db=False, **mel_kwargs):
         assert not all([amplitude_to_db, power_to_db]), "Choose amplitude_to_db or power_to_db, not both!"
-        self.mel_kwargs = kwargs
+        self.mel_kwargs = mel_kwargs
         self.amplitude_to_db = amplitude_to_db
         self.power_to_db = power_to_db
 

audio_toolset/mel_augmentation.py

@@ -59,9 +59,9 @@ class ShiftTime(object):
             # Set to silence for heading/ tailing
             shift = int(shift)
             if shift > 0:
-                augmented_data[:, :shift] = 0
+                augmented_data[:shift, :] = 0
             else:
-                augmented_data[:, shift:] = 0
+                augmented_data[shift:, :] = 0
             return augmented_data
         else:
             return x

audio_toolset/mel_transforms.py

@@ -0,0 +1,21 @@
+from typing import Union
+
+import numpy as np
+
+
+class Normalize(object):
+
+    def __init__(self, min_db_level: Union[int, float]):
+        self.min_db_level = min_db_level
+
+    def __call__(self, s: np.ndarray) -> np.ndarray:
+        return np.clip((s - self.min_db_level) / -self.min_db_level, 0, 1)
+
+
+class DeNormalize(object):
+
+    def __init__(self, min_db_level: Union[int, float]):
+        self.min_db_level = min_db_level
+
+    def __call__(self, s: np.ndarray) -> np.ndarray:
+        return (np.clip(s, 0, 1) * -self.min_db_level) + self.min_db_level