Merge pull request #207 from TensorSpeech/fix-pqmf

🔧 Fix pqmf

Author: dathudeptrai

tensorflow_tts/configs/mb_melgan.py

@@ -22,10 +22,10 @@ class MultiBandMelGANGeneratorConfig(MelGANGeneratorConfig):
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
-        self.subbands = 4
-        self.taps = 62
-        self.cutoff_ratio = 0.15
-        self.beta = 9.0
+        self.subbands = kwargs.pop("subbands", 4)
+        self.taps = kwargs.pop("taps", 62)
+        self.cutoff_ratio = kwargs.pop("cutoff_ratio", 0.142)
+        self.beta = kwargs.pop("beta", 9.0)
 
 
 class MultiBandMelGANDiscriminatorConfig(MelGANDiscriminatorConfig):

tensorflow_tts/models/mb_melgan.py

@@ -82,7 +82,7 @@ def __init__(self, config, **kwargs):
                 * np.cos(
                     (2 * k + 1)
                     * (np.pi / (2 * subbands))
-                    * (np.arange(taps + 1) - ((taps - 1) / 2))
+                    * (np.arange(taps + 1) - (taps / 2))
                     + (-1) ** k * np.pi / 4
                 )
             )
@@ -92,7 +92,7 @@ def __init__(self, config, **kwargs):
                 * np.cos(
                     (2 * k + 1)
                     * (np.pi / (2 * subbands))
-                    * (np.arange(taps + 1) - ((taps - 1) / 2))
+                    * (np.arange(taps + 1) - (taps / 2))
                     - (-1) ** k * np.pi / 4
                 )
             )
@@ -101,7 +101,7 @@ def __init__(self, config, **kwargs):
         analysis_filter = np.expand_dims(h_analysis, 1)
         analysis_filter = np.transpose(analysis_filter, (2, 1, 0))
 
-        synthesis_filter = np.expand_dims(h_analysis, 0)
+        synthesis_filter = np.expand_dims(h_synthesis, 0)
         synthesis_filter = np.transpose(synthesis_filter, (2, 1, 0))
 
         # filter for downsampling & upsampling