diff --git a/fluid/DeepASR/decoder/post_latgen_faster_mapped.cc b/fluid/DeepASR/decoder/post_latgen_faster_mapped.cc
index 87791d513162c5e53ccd1eaca1ec5a2a79c2afa0..f83730ce51a78ae302107914194f4e5699b4afd0 100644
--- a/fluid/DeepASR/decoder/post_latgen_faster_mapped.cc
+++ b/fluid/DeepASR/decoder/post_latgen_faster_mapped.cc
@@ -26,6 +26,7 @@ Decoder::Decoder(std::string trans_model_in_filename,
                  std::string word_syms_filename,
                  std::string fst_in_filename,
                  std::string logprior_in_filename,
+                 size_t beam_size,
                  kaldi::BaseFloat acoustic_scale) {
   const char *usage =
       "Generate lattices using neural net model.\n"
@@ -51,7 +52,7 @@ Decoder::Decoder(std::string trans_model_in_filename,
 
   int argc = 2;
   char *argv[] = {(char *)"post-latgen-faster-mapped",
-                  (char *)("--beam=" + std::string("11")).c_str()};
+                  (char *)("--beam=" + std::to_string(beam_size)).c_str()};
 
   po.Read(argc, argv);
 
diff --git a/fluid/DeepASR/decoder/post_latgen_faster_mapped.h b/fluid/DeepASR/decoder/post_latgen_faster_mapped.h
index 0bbb93065afab35a7bfd77d1e28b494bcabb42f3..9c234b8681690b9f1e3d30b61ac3b97b7055887f 100644
--- a/fluid/DeepASR/decoder/post_latgen_faster_mapped.h
+++ b/fluid/DeepASR/decoder/post_latgen_faster_mapped.h
@@ -29,6 +29,7 @@ public:
           std::string word_syms_filename,
           std::string fst_in_filename,
           std::string logprior_in_filename,
+          size_t beam_size,
           kaldi::BaseFloat acoustic_scale);
   ~Decoder();
 
diff --git a/fluid/DeepASR/decoder/pybind.cc b/fluid/DeepASR/decoder/pybind.cc
index 151eae3b836bcc855e6bddb36c4ee682692d4a94..4a9b27d4cf862e5c1492875512fdeba3e95ecb15 100644
--- a/fluid/DeepASR/decoder/pybind.cc
+++ b/fluid/DeepASR/decoder/pybind.cc
@@ -27,6 +27,7 @@ PYBIND11_MODULE(post_latgen_faster_mapped, m) {
                     std::string,
                     std::string,
                     std::string,
+                    size_t,
                     kaldi::BaseFloat>())
       .def("decode_from_file",
            (void (Decoder::*)(std::string, size_t)) & Decoder::decode_from_file,
diff --git a/fluid/DeepASR/examples/aishell/infer_by_ckpt.sh b/fluid/DeepASR/examples/aishell/infer_by_ckpt.sh
index 990daee3756fc21bb18352a2ac3813e787a10380..60a48ba5daeb70fe46359caa09dc1301f2b158f2 100644
--- a/fluid/DeepASR/examples/aishell/infer_by_ckpt.sh
+++ b/fluid/DeepASR/examples/aishell/infer_by_ckpt.sh
@@ -4,10 +4,11 @@ export CUDA_VISIBLE_DEVICES=2,3,4,5
 python -u ../../infer_by_ckpt.py --batch_size 96  \
                         --checkpoint checkpoints/deep_asr.pass_20.checkpoint \
                         --infer_feature_lst data/test_feature.lst  \
-                        --mean_var data/aishell/global_mean_var \
+                        --mean_var data/global_mean_var \
                         --frame_dim 80  \
                         --class_num 3040 \
                         --num_threads 24  \
+                        --beam_size 11 \
                         --decode_to_path $decode_to_path \
                         --trans_model mapped_decoder_data/exp/tri5a/final.mdl \
                         --log_prior mapped_decoder_data/logprior \
diff --git a/fluid/DeepASR/infer_by_ckpt.py b/fluid/DeepASR/infer_by_ckpt.py
index 881b5ba225665c4ad7b00727760500625b35a55b..1e0fb15c6d6f05aa1e054b37333b0fa0cb5cd8d9 100644
--- a/fluid/DeepASR/infer_by_ckpt.py
+++ b/fluid/DeepASR/infer_by_ckpt.py
@@ -27,6 +27,11 @@ def parse_args():
         type=int,
         default=32,
         help='The sequence number of a batch data. (default: %(default)d)')
+    parser.add_argument(
+        '--beam_size',
+        type=int,
+        default=11,
+        help='The beam size for decoding. (default: %(default)d)')
     parser.add_argument(
         '--minimum_batch_size',
         type=int,
@@ -211,7 +216,7 @@ def infer_from_ckpt(args):
 
     # init decoder
     decoder = Decoder(args.trans_model, args.vocabulary, args.graphs,
-                      args.log_prior, args.acoustic_scale)
+                      args.log_prior, args.beam_size, args.acoustic_scale)
 
     ltrans = [
         trans_add_delta.TransAddDelta(2, 2),