diff --git a/.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc.yaml b/.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc.yaml
new file mode 100644
index 000000000..3d02994f5
--- /dev/null
+++ b/.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc.yaml
@@ -0,0 +1,73 @@
+name: export-nemo-speaker-verification-to-onnx
+
+on:
+  workflow_dispatch:
+
+concurrency:
+  group: export-nemo-fast-conformer-hybrid-transducer-ctc-to-onnx-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  export-nemo-fast-conformer-hybrid-transducer-ctc-to-onnx:
+    if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
+    name: export NeMo fast conformer
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [macos-latest]
+        python-version: ["3.10"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install NeMo
+        shell: bash
+        run: |
+          BRANCH='main'
+          pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[asr]
+          pip install onnxruntime
+          pip install kaldi-native-fbank
+          pip install soundfile librosa
+
+      - name: Run
+        shell: bash
+        run: |
+          cd scripts/nemo/fast-conformer-hybrid-transducer-ctc
+          ./run-ctc.sh
+
+          mv -v sherpa-onnx-nemo* ../../..
+
+      - name: Download test waves
+        shell: bash
+        run: |
+          mkdir test_wavs
+          pushd test_wavs
+          curl -SL -O https://hf-mirror.com/csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small/resolve/main/test_wavs/0.wav
+          curl -SL -O https://hf-mirror.com/csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small/resolve/main/test_wavs/1.wav
+          curl -SL -O https://hf-mirror.com/csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small/resolve/main/test_wavs/8k.wav
+          curl -SL -O https://hf-mirror.com/csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small/resolve/main/test_wavs/trans.txt
+          popd
+
+          cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-ctc-80ms
+          cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-ctc-480ms
+          cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-ctc-1040ms
+
+          tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-ctc-80ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-ctc-80ms
+          tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-ctc-480ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-ctc-480ms
+          tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-ctc-1040ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-ctc-1040ms
+
+      - name: Release
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.tar.bz2
+          overwrite: true
+          repo_name: k2-fsa/sherpa-onnx
+          repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
+          tag: asr-models
diff --git a/scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md
new file mode 100644
index 000000000..bbb0f861f
--- /dev/null
+++ b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md
@@ -0,0 +1,9 @@
+# Introduction
+
+This folder contains scripts for exporting models from
+
+  - https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_en_fastconformer_hybrid_large_streaming_80ms
+  - https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_en_fastconformer_hybrid_large_streaming_480ms
+  - https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_en_fastconformer_hybrid_large_streaming_1040ms
+
+to `sherpa-onnx`.
diff --git a/scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc.py b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc.py
new file mode 100755
index 000000000..5751c90b5
--- /dev/null
+++ b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc.py
@@ -0,0 +1,117 @@
+#!/usr/bin/env python3
+import argparse
+from typing import Dict
+
+import nemo.collections.asr as nemo_asr
+import onnx
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        choices=["80", "480", "1040"],
+    )
+    return parser.parse_args()
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    while len(model.metadata_props):
+        model.metadata_props.pop()
+
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = str(value)
+
+    onnx.save(model, filename)
+
+
+@torch.no_grad()
+def main():
+    args = get_args()
+    model_name = f"stt_en_fastconformer_hybrid_large_streaming_{args.model}ms"
+
+    asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name=model_name)
+
+    with open("./tokens.txt", "w", encoding="utf-8") as f:
+        for i, s in enumerate(asr_model.joint.vocabulary):
+            f.write(f"{s} {i}\n")
+        f.write(f"<blk> {i+1}\n")
+        print("Saved to tokens.txt")
+
+    decoder_type = "ctc"
+    asr_model.change_decoding_strategy(decoder_type=decoder_type)
+    asr_model.eval()
+
+    assert asr_model.encoder.streaming_cfg is not None
+    if isinstance(asr_model.encoder.streaming_cfg.chunk_size, list):
+        chunk_size = asr_model.encoder.streaming_cfg.chunk_size[1]
+    else:
+        chunk_size = asr_model.encoder.streaming_cfg.chunk_size
+
+    if isinstance(asr_model.encoder.streaming_cfg.pre_encode_cache_size, list):
+        pre_encode_cache_size = asr_model.encoder.streaming_cfg.pre_encode_cache_size[1]
+    else:
+        pre_encode_cache_size = asr_model.encoder.streaming_cfg.pre_encode_cache_size
+    window_size = chunk_size + pre_encode_cache_size
+
+    print("chunk_size", chunk_size)
+    print("pre_encode_cache_size", pre_encode_cache_size)
+    print("window_size", window_size)
+
+    chunk_shift = chunk_size
+
+    # cache_last_channel: (batch_size, dim1, dim2, dim3)
+    cache_last_channel_dim1 = len(asr_model.encoder.layers)
+    cache_last_channel_dim2 = asr_model.encoder.streaming_cfg.last_channel_cache_size
+    cache_last_channel_dim3 = asr_model.encoder.d_model
+
+    # cache_last_time: (batch_size, dim1, dim2, dim3)
+    cache_last_time_dim1 = len(asr_model.encoder.layers)
+    cache_last_time_dim2 = asr_model.encoder.d_model
+    cache_last_time_dim3 = asr_model.encoder.conv_context_size[0]
+
+    asr_model.set_export_config({"decoder_type": "ctc", "cache_support": True})
+
+    filename = "model.onnx"
+
+    asr_model.export(filename)
+
+    meta_data = {
+        "vocab_size": asr_model.decoder.vocab_size,
+        "window_size": window_size,
+        "chunk_shift": chunk_shift,
+        "normalize_type": "None",
+        "cache_last_channel_dim1": cache_last_channel_dim1,
+        "cache_last_channel_dim2": cache_last_channel_dim2,
+        "cache_last_channel_dim3": cache_last_channel_dim3,
+        "cache_last_time_dim1": cache_last_time_dim1,
+        "cache_last_time_dim2": cache_last_time_dim2,
+        "cache_last_time_dim3": cache_last_time_dim3,
+        "subsampling_factor": 8,
+        "model_type": "EncDecHybridRNNTCTCBPEModel",
+        "version": "1",
+        "model_author": "NeMo",
+        "url": f"https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/{model_name}",
+        "comment": "Only the CTC branch is exported",
+    }
+    add_meta_data(filename, meta_data)
+
+    print(meta_data)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc.sh b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc.sh
new file mode 100755
index 000000000..6fae9b949
--- /dev/null
+++ b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc.sh
@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+
+set -ex
+
+if [ ! -e ./0.wav ]; then
+  # curl -SL -O https://hf-mirror.com/csukuangfj/icefall-asr-librispeech-streaming-zipformer-small-2024-03-18/resolve/main/test_wavs/0.wav
+  curl -SL -O https://huggingface.co/csukuangfj/icefall-asr-librispeech-streaming-zipformer-small-2024-03-18/resolve/main/test_wavs/0.wav
+fi
+
+ms=(
+80
+480
+1040
+)
+
+for m in ${ms[@]}; do
+  ./export-onnx-ctc.py --model $m
+  d=sherpa-onnx-nemo-streaming-fast-conformer-ctc-${m}ms
+  if [ ! -f $d/model.onnx ]; then
+    mkdir -p $d
+    mv -v model.onnx $d/
+    mv -v tokens.txt $d/
+    ls -lh $d
+  fi
+done
+
+# Now test the exported models
+
+for m in ${ms[@]}; do
+  d=sherpa-onnx-nemo-streaming-fast-conformer-ctc-${m}ms
+  python3 ./test-onnx-ctc.py \
+    --model $d/model.onnx \
+    --tokens $d/tokens.txt \
+    --wav ./0.wav
+done
diff --git a/scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-ctc.py b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-ctc.py
new file mode 100755
index 000000000..77c7a526b
--- /dev/null
+++ b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-ctc.py
@@ -0,0 +1,197 @@
+#!/usr/bin/env python3
+
+import argparse
+from pathlib import Path
+
+import kaldi_native_fbank as knf
+import numpy as np
+import onnxruntime as ort
+import torch
+import soundfile as sf
+import librosa
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", type=str, required=True, help="Path to model.onnx")
+
+    parser.add_argument("--tokens", type=str, required=True, help="Path to tokens.txt")
+
+    parser.add_argument("--wav", type=str, required=True, help="Path to test.wav")
+
+    return parser.parse_args()
+
+
+def create_fbank():
+    opts = knf.FbankOptions()
+    opts.frame_opts.dither = 0
+    opts.frame_opts.remove_dc_offset = False
+    opts.frame_opts.window_type = "hann"
+
+    opts.mel_opts.low_freq = 0
+    opts.mel_opts.num_bins = 80
+
+    opts.mel_opts.is_librosa = True
+
+    fbank = knf.OnlineFbank(opts)
+    return fbank
+
+
+def compute_features(audio, fbank):
+    assert len(audio.shape) == 1, audio.shape
+    fbank.accept_waveform(16000, audio)
+    ans = []
+    processed = 0
+    while processed < fbank.num_frames_ready:
+        ans.append(np.array(fbank.get_frame(processed)))
+        processed += 1
+    ans = np.stack(ans)
+    return ans
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.model = ort.InferenceSession(
+            filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+        self.window_size = int(meta["window_size"])
+        self.chunk_shift = int(meta["chunk_shift"])
+
+        self.cache_last_channel_dim1 = int(meta["cache_last_channel_dim1"])
+        self.cache_last_channel_dim2 = int(meta["cache_last_channel_dim2"])
+        self.cache_last_channel_dim3 = int(meta["cache_last_channel_dim3"])
+
+        self.cache_last_time_dim1 = int(meta["cache_last_time_dim1"])
+        self.cache_last_time_dim2 = int(meta["cache_last_time_dim2"])
+        self.cache_last_time_dim3 = int(meta["cache_last_time_dim3"])
+
+        self.init_cache_state()
+
+    def init_cache_state(self):
+        self.cache_last_channel = torch.zeros(
+            1,
+            self.cache_last_channel_dim1,
+            self.cache_last_channel_dim2,
+            self.cache_last_channel_dim3,
+            dtype=torch.float32,
+        ).numpy()
+
+        self.cache_last_time = torch.zeros(
+            1,
+            self.cache_last_time_dim1,
+            self.cache_last_time_dim2,
+            self.cache_last_time_dim3,
+            dtype=torch.float32,
+        ).numpy()
+
+        self.cache_last_channel_len = torch.ones([1], dtype=torch.int64).numpy()
+
+    def __call__(self, x: np.ndarray):
+        # x: (T, C)
+        x = torch.from_numpy(x)
+        x = x.t().unsqueeze(0)
+        # x: [1, C, T]
+        x_lens = torch.tensor([x.shape[-1]], dtype=torch.int64)
+
+        (
+            log_probs,
+            log_probs_len,
+            cache_last_channel_next,
+            cache_last_time_next,
+            cache_last_channel_len_next,
+        ) = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+                self.model.get_outputs()[2].name,
+                self.model.get_outputs()[3].name,
+                self.model.get_outputs()[4].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+                self.model.get_inputs()[2].name: self.cache_last_channel,
+                self.model.get_inputs()[3].name: self.cache_last_time,
+                self.model.get_inputs()[4].name: self.cache_last_channel_len,
+            },
+        )
+        self.cache_last_channel = cache_last_channel_next
+        self.cache_last_time = cache_last_time_next
+        self.cache_last_channel_len = cache_last_channel_len_next
+
+        # [T, vocab_size]
+        return torch.from_numpy(log_probs).squeeze(0)
+
+
+def main():
+    args = get_args()
+    assert Path(args.model).is_file(), args.model
+    assert Path(args.tokens).is_file(), args.tokens
+    assert Path(args.wav).is_file(), args.wav
+
+    print(vars(args))
+
+    model = OnnxModel(args.model)
+
+    id2token = dict()
+    with open(args.tokens, encoding="utf-8") as f:
+        for line in f:
+            t, idx = line.split()
+            id2token[int(idx)] = t
+
+    fbank = create_fbank()
+    audio, sample_rate = sf.read(args.wav, dtype="float32", always_2d=True)
+    audio = audio[:, 0]  # only use the first channel
+    if sample_rate != 16000:
+        audio = librosa.resample(
+            audio,
+            orig_sr=sample_rate,
+            target_sr=16000,
+        )
+        sample_rate = 16000
+
+    window_size = model.window_size
+    chunk_shift = model.chunk_shift
+
+    blank = len(id2token) - 1
+    prev = -1
+    ans = []
+
+    features = compute_features(audio, fbank)
+    num_chunks = (features.shape[0] - window_size) // chunk_shift + 1
+    for i in range(num_chunks):
+        start = i * chunk_shift
+        end = start + window_size
+        chunk = features[start:end, :]
+
+        log_probs = model(chunk)
+        ids = torch.argmax(log_probs, dim=1).tolist()
+        for i in ids:
+            if i != blank and i != prev:
+                ans.append(i)
+            prev = i
+
+    tokens = [id2token[i] for i in ans]
+    underline = "▁"
+    #  underline = b"\xe2\x96\x81".decode()
+    text = "".join(tokens).replace(underline, " ").strip()
+    print(args.wav)
+    print(text)
+
+
+main()