diff --git a/dnn/torch/fargan/dataset.py b/dnn/torch/fargan/dataset.py
index 0f2e4a819..2dfbb0b56 100644
--- a/dnn/torch/fargan/dataset.py
+++ b/dnn/torch/fargan/dataset.py
@@ -36,7 +36,7 @@ def __init__(self,
         self.features = np.lib.stride_tricks.as_strided(self.features, shape=(self.nb_sequences, self.sequence_length*2+4, nb_features),
                                            strides=(self.sequence_length*self.nb_features*sizeof, self.nb_features*sizeof, sizeof))
         #self.periods = np.round(50*self.features[:,:,self.nb_used_features-2]+100).astype('int')
-        self.periods = np.round(256./2**(self.features[:,:,self.nb_used_features-2]+1.5)).astype('int')
+        self.periods = np.round(np.clip(256./2**(self.features[:,:,self.nb_used_features-2]+1.5), 32, 255)).astype('int')
 
         self.lpc = self.features[:, :, self.nb_used_features:]
         self.features = self.features[:, :, :self.nb_used_features]
diff --git a/dnn/torch/fargan/fargan.py b/dnn/torch/fargan/fargan.py
index 84053fd08..39b5ddaf7 100644
--- a/dnn/torch/fargan/fargan.py
+++ b/dnn/torch/fargan/fargan.py
@@ -157,7 +157,7 @@ def __init__(self, feature_dim=20, cond_size=256, pembed_dims=12):
         self.feature_dim = feature_dim
         self.cond_size = cond_size
 
-        self.pembed = nn.Embedding(225, pembed_dims)
+        self.pembed = nn.Embedding(224, pembed_dims)
         self.fdense1 = nn.Linear(self.feature_dim + pembed_dims, 64, bias=False)
         self.fconv1 = nn.Conv1d(64, 128, kernel_size=3, padding='valid', bias=False)
         self.fconv2 = nn.Conv1d(128, 80*4, kernel_size=3, padding='valid', bias=False)
diff --git a/dnn/torch/fargan/test_fargan.py b/dnn/torch/fargan/test_fargan.py
index eb12aca8f..d47400f86 100644
--- a/dnn/torch/fargan/test_fargan.py
+++ b/dnn/torch/fargan/test_fargan.py
@@ -49,7 +49,7 @@
 lpc = features[:,4-1:-1,nb_used_features:]
 features = features[:, :, :nb_used_features]
 #periods = np.round(50*features[:,:,nb_used_features-2]+100).astype('int')
-periods = np.round(256./2**(features[:,:,nb_used_features-2]+1.5)).astype('int')
+periods = np.round(np.clip(256./2**(features[:,:,nb_used_features-2]+1.5), 32, 255)).astype('int')
 
 
 nb_frames = features.shape[1]