diff --git a/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp b/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp
index 54f9872a7..37ea317bb 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/.transformer.py.swp b/egs/tedlium2/ASR/conformer_ctc3/.transformer.py.swp
index 53fadf4d3..a052c8f62 100644
Binary files a/egs/tedlium2/ASR/conformer_ctc3/.transformer.py.swp and b/egs/tedlium2/ASR/conformer_ctc3/.transformer.py.swp differ
diff --git a/egs/tedlium2/ASR/conformer_ctc3/train.py b/egs/tedlium2/ASR/conformer_ctc3/train.py
index dc4b0348d..74490021a 100755
--- a/egs/tedlium2/ASR/conformer_ctc3/train.py
+++ b/egs/tedlium2/ASR/conformer_ctc3/train.py
@@ -638,6 +638,7 @@ def compute_loss(
             ctc_loss = (1-params.interctc_weight) * ctc_loss + params.interctc_weight * inter_ctc_loss
 
         else:
+            '''
             dense_fsa_vec = k2.DenseFsaVec(
                 nnet_output,
                 supervision_segments,
@@ -651,6 +652,60 @@ def compute_loss(
                 reduction=params.reduction,
                 use_double_scores=params.use_double_scores,
             )
+            '''
+            dense_fsa_vec = k2.DenseFsaVec(
+                nnet_output[0],
+                supervision_segments,
+                allow_truncate=params.subsampling_factor - 1,
+            )
+            
+            dense_fsa_vec_inter = [
+                k2.DenseFsaVec(
+                    nnet_output[1][2],
+                    supervision_segments,
+                    allow_truncate=params.subsampling_factor - 1,
+                ),
+                k2.DenseFsaVec(
+                    nnet_output[1][5],
+                    supervision_segments,
+                    allow_truncate=params.subsampling_factor - 1,
+                ),
+                k2.DenseFsaVec(
+                    nnet_output[1][8],
+                    supervision_segments,
+                    allow_truncate=params.subsampling_factor - 1,
+                ),
+                k2.DenseFsaVec(
+                    nnet_output[1][11],
+                    supervision_segments,
+                    allow_truncate=params.subsampling_factor - 1,
+                ),
+                k2.DenseFsaVec(
+                    nnet_output[1][14],
+                    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,
+            ) 
+            
+            inter_ctc_loss = 0
+            for fsa in dense_fsa_vec_inter:
+                inter_ctc_loss += k2.ctc_loss(
+                    decoding_graph=decoding_graph,
+                    dense_fsa_vec=fsa,
+                    output_beam=params.beam_size,
+                    reduction=params.reduction,
+                    use_double_scores=params.use_double_scores,
+                ) 
+            
+            ctc_loss = (1-params.interctc_weight) * ctc_loss + params.interctc_weight * inter_ctc_loss
 
         if params.att_rate > 0.0:
             with torch.set_grad_enabled(is_training):