diff --git a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py
index 5d0c9ea430..e5ceb36831 100755
--- a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py
+++ b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py
@@ -48,7 +48,8 @@
   --joiner-dim 512 \
   --causal True \
   --chunk-size 16 \
-  --left-context-frames 128
+  --left-context-frames 128 \
+  --fp16 True
 
 The --chunk-size in training is "16,32,64,-1", so we select one of them
 (excluding -1) during streaming export. The same applies to `--left-context`,
@@ -73,6 +74,7 @@
 import torch
 import torch.nn as nn
 from decoder import Decoder
+from onnxconverter_common import float16
 from onnxruntime.quantization import QuantType, quantize_dynamic
 from scaling_converter import convert_scaled_to_non_scaled
 from train import add_model_arguments, get_model, get_params
@@ -154,6 +156,13 @@ def get_parser():
         help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
     )
 
+    parser.add_argument(
+        "--fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to export models in fp16",
+    )
+
     add_model_arguments(parser)
 
     return parser
@@ -479,7 +488,6 @@ def build_inputs_outputs(tensors, i):
 
     add_meta_data(filename=encoder_filename, meta_data=meta_data)
 
-
 def export_decoder_model_onnx(
     decoder_model: OnnxDecoder,
     decoder_filename: str,
@@ -747,11 +755,29 @@ def main():
     )
     logging.info(f"Exported joiner to {joiner_filename}")
 
+    if(params.fp16) :
+        logging.info("Generate fp16 models")
+
+        encoder = onnx.load(encoder_filename)
+        encoder_fp16 = float16.convert_float_to_float16(encoder, keep_io_types=True)
+        encoder_filename_fp16 = params.exp_dir / f"encoder-{suffix}.fp16.onnx"
+        onnx.save(encoder_fp16,encoder_filename_fp16)
+
+        decoder = onnx.load(decoder_filename)
+        decoder_fp16 = float16.convert_float_to_float16(decoder, keep_io_types=True)
+        decoder_filename_fp16 = params.exp_dir / f"decoder-{suffix}.fp16.onnx"
+        onnx.save(decoder_fp16,decoder_filename_fp16)
+
+        joiner = onnx.load(joiner_filename)
+        joiner_fp16 = float16.convert_float_to_float16(joiner, keep_io_types=True)
+        joiner_filename_fp16 = params.exp_dir / f"joiner-{suffix}.fp16.onnx"
+        onnx.save(joiner_fp16,joiner_filename_fp16)
+
     # Generate int8 quantization models
     # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
 
     logging.info("Generate int8 quantization models")
-
+    
     encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
     quantize_dynamic(
         model_input=encoder_filename,
diff --git a/egs/librispeech/ASR/zipformer/export-onnx.py b/egs/librispeech/ASR/zipformer/export-onnx.py
index 3682f0b625..ed8a0ef0fd 100755
--- a/egs/librispeech/ASR/zipformer/export-onnx.py
+++ b/egs/librispeech/ASR/zipformer/export-onnx.py
@@ -48,8 +48,8 @@
   --joiner-dim 512 \
   --causal False \
   --chunk-size "16,32,64,-1" \
-  --left-context-frames "64,128,256,-1"
-
+  --left-context-frames "64,128,256,-1" \
+  --fp16 True
 It will generate the following 3 files inside $repo/exp:
 
   - encoder-epoch-99-avg-1.onnx
@@ -70,6 +70,7 @@
 import torch
 import torch.nn as nn
 from decoder import Decoder
+from onnxconverter_common import float16
 from onnxruntime.quantization import QuantType, quantize_dynamic
 from scaling_converter import convert_scaled_to_non_scaled
 from train import add_model_arguments, get_model, get_params
@@ -151,6 +152,13 @@ def get_parser():
         help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
     )
 
+    parser.add_argument(
+        "--fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to export models in fp16",
+    )
+
     add_model_arguments(parser)
 
     return parser
@@ -584,6 +592,24 @@ def main():
     )
     logging.info(f"Exported joiner to {joiner_filename}")
 
+    if(params.fp16) :
+        logging.info("Generate fp16 models")
+
+        encoder = onnx.load(encoder_filename)
+        encoder_fp16 = float16.convert_float_to_float16(encoder, keep_io_types=True)
+        encoder_filename_fp16 = params.exp_dir / f"encoder-{suffix}.fp16.onnx"
+        onnx.save(encoder_fp16,encoder_filename_fp16)
+
+        decoder = onnx.load(decoder_filename)
+        decoder_fp16 = float16.convert_float_to_float16(decoder, keep_io_types=True)
+        decoder_filename_fp16 = params.exp_dir / f"decoder-{suffix}.fp16.onnx"
+        onnx.save(decoder_fp16,decoder_filename_fp16)
+
+        joiner = onnx.load(joiner_filename)
+        joiner_fp16 = float16.convert_float_to_float16(joiner, keep_io_types=True)
+        joiner_filename_fp16 = params.exp_dir / f"joiner-{suffix}.fp16.onnx"
+        onnx.save(joiner_fp16,joiner_filename_fp16)
+
     # Generate int8 quantization models
     # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
 
diff --git a/requirements.txt b/requirements.txt
index 226adaba1b..d97263142c 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -12,6 +12,7 @@ onnx>=1.15.0
 onnxruntime>=1.16.3
 onnxoptimizer
 onnxsim
+onnxconverter_common
 
 # style check session:
 black==22.3.0