diff --git a/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp b/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp
index 38d6a2df6..e29f93389 100644
Binary files a/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp and b/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp differ
diff --git a/egs/tedlium2/ASR/conformer_ctc3/train.py b/egs/tedlium2/ASR/conformer_ctc3/train.py
index ffbe41c7c..0557a7c20 100755
--- a/egs/tedlium2/ASR/conformer_ctc3/train.py
+++ b/egs/tedlium2/ASR/conformer_ctc3/train.py
@@ -543,20 +543,36 @@ def compute_loss(
         #token_ids = convert_texts_into_ids(texts, graph_compiler.sp)
         token_ids = graph_compiler.texts_to_ids(texts)
         decoding_graph = graph_compiler.compile(token_ids)
+        
+        if params.interctc:
+            dense_fsa_vec1 = k2.DenseFsaVec(
+                nnet_output[8],
+                supervision_segments,
+                allow_truncate=params.subsampling_factor - 1,
+            )
 
-        dense_fsa_vec = k2.DenseFsaVec(
-            nnet_output,
-            supervision_segments,
-            allow_truncate=params.subsampling_factor - 1,
-        )
+            ctc_loss = k2.ctc_loss(
+                decoding_graph=decoding_graph,
+                dense_fsa_vec=dense_fsa_vec,
+                output_beam=params.beam_size,
+                reduction=params.reduction,
+                use_double_scores=params.use_double_scores,
+            )
 
-        ctc_loss = k2.ctc_loss(
-            decoding_graph=decoding_graph,
-            dense_fsa_vec=dense_fsa_vec,
-            output_beam=params.beam_size,
-            reduction=params.reduction,
-            use_double_scores=params.use_double_scores,
-        )
+        else:
+            dense_fsa_vec = k2.DenseFsaVec(
+                nnet_output,
+                supervision_segments,
+                allow_truncate=params.subsampling_factor - 1,
+            )
+
+            ctc_loss = k2.ctc_loss(
+                decoding_graph=decoding_graph,
+                dense_fsa_vec=dense_fsa_vec,
+                output_beam=params.beam_size,
+                reduction=params.reduction,
+                use_double_scores=params.use_double_scores,
+            )
 
         if params.att_rate > 0.0:
             with torch.set_grad_enabled(is_training):