diff --git a/layers/common_layers.py b/layers/common_layers.py
index b6f72bc1..f7b8e7ed 100644
--- a/layers/common_layers.py
+++ b/layers/common_layers.py
@@ -208,7 +208,7 @@ class Attention(nn.Module):
             _, n = prev_alpha.max(1)
             val, n2 = alpha.max(1)
             for b in range(alignment.shape[0]):
-                alpha[b, n[b] + 2:] = 0
+                alpha[b, n[b] + 3:] = 0
                 alpha[b, :(n[b] - 1)] = 0  # ignore all previous states to prevent repetition.
                 alpha[b, (n[b] - 2)] = 0.01 * val[b]  # smoothing factor for the prev step
         # compute attention weights
diff --git a/utils/audio.py b/utils/audio.py
index fb3edad3..e985dbf2 100644
--- a/utils/audio.py
+++ b/utils/audio.py
@@ -230,12 +230,13 @@ class AudioProcessor(object):
         x = np.sign(wav) / mu * ((1 + mu) ** np.abs(wav) - 1)
         return x
 
-    def load_wav(self, filename, encode=False):
-        x, sr = sf.read(filename)
-        # x, sr = librosa.load(filename, sr=self.sample_rate)
+    def load_wav(self, filename, sr=None):
+        if sr is None:
+            x, sr = sf.read(filename)
+        else:
+            x, sr = librosa.load(filename, sr=sr)
         if self.do_trim_silence:
             x = self.trim_silence(x)
-        # sr, x = io.wavfile.read(filename)
         assert self.sample_rate == sr, "Expected sampling rate {} but file " \
                                        "{} has {}.".format(self.sample_rate,
                                                            filename,