diff --git a/decoders/swig_wrapper.py b/decoders/swig_wrapper.py
index 0a0579ad0..3051f4e82 100644
--- a/decoders/swig_wrapper.py
+++ b/decoders/swig_wrapper.py
@@ -46,7 +46,7 @@ def ctc_greedy_decoder(probs_seq, vocabulary):
     :rtype: str
     """
     result = swig_decoders.ctc_greedy_decoder(probs_seq.tolist(), vocabulary)
-    return result.decode('utf-8')
+    return result
 
 
 def ctc_beam_search_decoder(probs_seq,
diff --git a/model_utils/network.py b/model_utils/network.py
index c8fa95dc4..1e7545ee6 100644
--- a/model_utils/network.py
+++ b/model_utils/network.py
@@ -686,8 +686,9 @@ class DeepSpeech2(nn.Layer):
     def decode(self, audio, audio_len, vocab_list, decoding_method,
                lang_model_path, beam_alpha, beam_beta, beam_size, cutoff_prob,
                cutoff_top_n, num_processes):
-        _, probs, _ = self.predict(audio, audio_len)
-        return self.decode_probs(probs.numpy(), vocab_list, decoding_method,
+        _, probs, audio_lens = self.predict(audio, audio_len)
+        probs_split = [probs[i, :l, :] for i, l in enumerate(audio_lens)]
+        return self.decode_probs(probs_split, vocab_list, decoding_method,
                                  lang_model_path, beam_alpha, beam_beta,
                                  beam_size, cutoff_prob, cutoff_top_n,
                                  num_processes)