diff --git a/.github/workflows/export-sense-voice-to-ascend-npu.yaml b/.github/workflows/export-sense-voice-to-ascend-npu.yaml
new file mode 100644
index 0000000000..bdabfdfd5f
--- /dev/null
+++ b/.github/workflows/export-sense-voice-to-ascend-npu.yaml
@@ -0,0 +1,216 @@
+name: export-sense-voice-to-ascend-npu
+
+on:
+  push:
+    branches:
+      - export-sense-voice-ascend
+  workflow_dispatch:
+
+concurrency:
+  group: export-sense-voice-to-ascend-npu-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  export-sense-voice-to-ascend-npu:
+    if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
+    name: ${{ matrix.framework }}
+    runs-on: ${{ matrix.os }}
+    container:
+      # image: ascendai/cann:latest
+      # image: ascendai/cann:8.1.rc1-910b-ubuntu22.04-py3.10
+      # see https://hub.docker.com/r/gpustack/ascendai-cann/tags?name=8.0
+      image: gpustack/ascendai-cann:8.0.RC3-910b-ubuntu20.04-py3.9
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.8"]
+        framework: ["FunASR", "WSYue-ASR"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Show Python
+        shell: bash
+        run: |
+          python3 --version
+
+      - name: Install curl
+        shell: bash
+        run: apt-get update && apt-get install -y curl bzip2
+
+      - name: Verify environment
+        shell: bash
+        run: |
+          ls -lh /usr/local/Ascend/ascend-toolkit/set_env.sh
+
+          find /usr/local/Ascend -name "libascend*.so" 2>/dev/null
+
+
+          source /usr/local/Ascend/ascend-toolkit/set_env.sh
+          export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/devlib/linux/x86_64:$LD_LIBRARY_PATH
+
+          echo "CANN environment:"
+          which atc || echo "atc not found"
+          atc --help
+
+      - name: Install Python dependencies
+        shell: bash
+        run: |
+          python3 -m pip install "numpy<2" \
+                  onnx==1.17.0 \
+                  torch==2.0.0+cpu -f https://download.pytorch.org/whl/torch \
+                  attrs psutil scipy decorator cloudpickle ml-dtypes tornado \
+                  sentencepiece \
+                  pyyaml
+
+      - name: Run SenseVoice from FunAsr
+        if: matrix.framework == 'FunASR'
+        shell: bash
+        run: |
+          cd scripts/sense-voice/ascend-npu
+
+          curl -SL -O https://hf-mirror.com/FunAudioLLM/SenseVoiceSmall/resolve/main/am.mvn
+          curl -SL -O https://hf-mirror.com/FunAudioLLM/SenseVoiceSmall/resolve/main/model.pt
+          curl -SL -O https://hf-mirror.com/FunAudioLLM/SenseVoiceSmall/resolve/main/chn_jpn_yue_eng_ko_spectok.bpe.model
+
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/en.wav
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/ja.wav
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/ko.wav
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/yue.wav
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/zh.wav
+
+          rm -f README.md || true
+
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/README.md
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/LICENSE
+
+          echo "export to onnx"
+
+          python3 ./export_onnx.py
+
+          ls -lh *.onnx
+
+          source /usr/local/Ascend/ascend-toolkit/set_env.sh
+          export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/devlib/linux/x86_64:$LD_LIBRARY_PATH
+
+          atc --model=./model.onnx \
+            --framework=5 \
+            --host_env_os=linux \
+            --host_env_cpu=aarch64 \
+            --output=model \
+            --input_format=ND \
+            --input_shape="x:1,-1,560;prompt:4" \
+            --soc_version="Ascend910B"
+
+          ls -lh *.om
+
+          echo "collect results"
+          d=sherpa-onnx-ascend-910B-sense-voice-zh-en-ja-ko-yue-2024-07-17
+
+          mkdir -p $d
+          mkdir -p $d/test_wavs
+
+          cp -v README.md $d
+          cp -v LICENSE $d
+          cp -v model_linux_aarch64.om $d/model.om
+          cp -v tokens.txt $d
+          cp -v test_om.py $d
+          cp -v *.wav $d/test_wavs
+          ls -lh $d
+          tar cjfv $d.tar.bz2 $d
+          ls -lh *.tar.bz2
+          rm -rf $d
+
+          echo "----show---"
+          ls -lh *.tar.bz2
+
+          mv *.tar.bz2 ../../..
+
+      - name: Run SenseVoice from WSYue-ASR
+        if: matrix.framework == 'WSYue-ASR'
+        shell: bash
+        run: |
+          cd scripts/sense-voice/ascend-npu
+
+          curl -SL -O https://huggingface.co/ASLP-lab/WSYue-ASR/resolve/main/sensevoice_small_yue/model.pt
+
+          curl -SL -O https://hf-mirror.com/FunAudioLLM/SenseVoiceSmall/resolve/main/am.mvn
+          curl -SL -O https://hf-mirror.com/FunAudioLLM/SenseVoiceSmall/resolve/main/chn_jpn_yue_eng_ko_spectok.bpe.model
+
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/en.wav
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/yue.wav
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/test_wavs/zh.wav
+
+          for i in $(seq 0 17); do
+            curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09/resolve/main/test_wavs/yue-$i.wav
+          done
+
+          rm -f README.md || true
+
+          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09/resolve/main/README.md
+
+          echo "export to onnx"
+          python3 ./export_onnx.py
+
+          ls -lh *.onnx
+
+          source /usr/local/Ascend/ascend-toolkit/set_env.sh
+          export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/devlib/linux/x86_64:$LD_LIBRARY_PATH
+
+          atc --model=./model.onnx \
+            --framework=5 \
+            --host_env_os=linux \
+            --host_env_cpu=aarch64 \
+            --output=model \
+            --input_format=ND \
+            --input_shape="x:1,-1,560;prompt:4" \
+            --soc_version="Ascend910B"
+
+          ls -lh *.om
+
+          echo "collect results"
+          d=sherpa-onnx-ascend-910B-sense-voice-zh-en-ja-ko-yue-2025-09-09
+
+          mkdir -p $d
+          mkdir -p $d/test_wavs
+
+          cp -v README.md $d
+          cp -v model_linux_aarch64.om $d/model.om
+          cp -v tokens.txt $d
+          cp -v test_om.py $d
+          cp -v *.wav $d/test_wavs
+          ls -lh $d
+          tar cjfv $d.tar.bz2 $d
+          ls -lh *.tar.bz2
+          rm -rf $d
+
+          echo "----show---"
+          ls -lh *.tar.bz2
+
+          mv *.tar.bz2 ../../..
+
+      - name: Release
+        if: github.repository_owner == 'csukuangfj'
+        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
+
+      - name: Release
+        if: github.repository_owner == 'k2-fsa'
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.tar.bz2
+          overwrite: true
+          tag: asr-models
diff --git a/scripts/sense-voice/ascend-npu/export_onnx.py b/scripts/sense-voice/ascend-npu/export_onnx.py
new file mode 100755
index 0000000000..596b50b397
--- /dev/null
+++ b/scripts/sense-voice/ascend-npu/export_onnx.py
@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+# Copyright      2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from typing import List, Tuple
+
+import sentencepiece as spm
+import torch
+
+from torch_model import SenseVoiceSmall
+
+
+def load_cmvn(filename) -> Tuple[List[float], List[float]]:
+    neg_mean = None
+    inv_stddev = None
+
+    with open(filename) as f:
+        for line in f:
+            if not line.startswith("<LearnRateCoef>"):
+                continue
+            t = line.split()[3:-1]
+
+            if neg_mean is None:
+                neg_mean = list(map(lambda x: float(x), t))
+            else:
+                inv_stddev = list(map(lambda x: float(x), t))
+
+    return neg_mean, inv_stddev
+
+
+def generate_tokens(sp):
+    with open("tokens.txt", "w", encoding="utf-8") as f:
+        for i in range(sp.vocab_size()):
+            f.write(f"{sp.id_to_piece(i)} {i}\n")
+    print("saved to tokens.txt")
+
+
+class ModelWrapper(torch.nn.Module):
+    def __init__(self, m):
+        super().__init__()
+        self.m = m
+
+    def forward(self, x, prompt):
+        logits = self.m(x[None], prompt)[0]
+        part1 = logits[:4]
+        part2 = logits[4:]
+        part1 = part1.reshape(4, 25055)
+        part2 = part2.reshape(x.size(0), 25055)
+        return part1, part2
+
+
+@torch.no_grad()
+def main():
+    sp = spm.SentencePieceProcessor()
+    sp.load("./chn_jpn_yue_eng_ko_spectok.bpe.model")
+    generate_tokens(sp)
+
+    print("loading model")
+
+    state_dict = torch.load("./model.pt", map_location="cpu")
+    if "state_dict" in state_dict:
+        state_dict = state_dict["state_dict"]
+
+    neg_mean, inv_stddev = load_cmvn("./am.mvn")
+
+    neg_mean = torch.tensor(neg_mean, dtype=torch.float32)
+    inv_stddev = torch.tensor(inv_stddev, dtype=torch.float32)
+
+    model = SenseVoiceSmall(neg_mean=neg_mean, inv_stddev=inv_stddev)
+    model.load_state_dict(state_dict)
+    model.eval()
+    del state_dict
+
+    model = ModelWrapper(model)
+    model.eval()
+
+    x = torch.randn(1, 93, 560, dtype=torch.float32)
+
+    language = 3
+    text_norm = 15
+    prompt = torch.tensor([language, 1, 2, text_norm], dtype=torch.int32)
+
+    opset_version = 14
+    filename = "model.onnx"
+    torch.onnx.export(
+        model.m,
+        (x, prompt),
+        filename,
+        opset_version=opset_version,
+        input_names=["x", "prompt"],
+        output_names=["logits"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "logits": {0: "N", 1: "T_4"},
+        },
+    )
+    print(f"saved to {filename}")
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20251018)
+    main()
diff --git a/scripts/sense-voice/ascend-npu/test_om.py b/scripts/sense-voice/ascend-npu/test_om.py
new file mode 100755
index 0000000000..9415768cd9
--- /dev/null
+++ b/scripts/sense-voice/ascend-npu/test_om.py
@@ -0,0 +1,166 @@
+#!/usr/bin/env python3
+# Copyright      2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import argparse
+from typing import Tuple
+
+import kaldi_native_fbank as knf
+import numpy as np
+import soundfile as sf
+import torch
+from ais_bench.infer.interface import InferSession
+
+
+class OmModel:
+    def __init__(self):
+        self.model = InferSession(device_id=0, model_path="./model.om", debug=False)
+
+        print("---model---")
+        for i in self.model.get_inputs():
+            print(i.name, i.datatype, i.shape)
+
+        print("-----")
+
+        for i in self.model.get_outputs():
+            print(i.name, i.datatype, i.shape)
+
+    def __call__(self, x, prompt=None, language=None, text_norm=None):
+        if language is not None and text_norm is not None:
+            x_len = np.array([x.shape[1]], dtype=np.int32)
+            print(x.shape, x_len, language, text_norm)
+            return self.model.infer(
+                [x, x_len, language, text_norm], mode="dymshape", custom_sizes=10000000
+            )[0][0]
+        else:
+            return self.model.infer(
+                [x, prompt], mode="dymshape", custom_sizes=10000000
+            )[0][0]
+
+        return self.model.infer([x, prompt], mode="static", custom_sizes=10000000)[0][0]
+        logits_part1, logits_part2 = self.model.infer(
+            [x, prompt], mode="dymshape", custom_sizes=10000000
+        )
+        logits = np.concatenate([logits_part1, logits_part2], axis=0)
+        return logits
+
+
+def load_audio(filename: str) -> Tuple[np.ndarray, int]:
+    data, sample_rate = sf.read(
+        filename,
+        always_2d=True,
+        dtype="float32",
+    )
+    data = data[:, 0]  # use only the first channel
+    samples = np.ascontiguousarray(data)
+    return samples, sample_rate
+
+
+def load_tokens(filename):
+    ans = dict()
+    i = 0
+    with open(filename, encoding="utf-8") as f:
+        for line in f:
+            ans[i] = line.strip().split()[0]
+            i += 1
+    return ans
+
+
+def compute_feat(
+    samples,
+    sample_rate,
+    window_size: int = 7,  # lfr_m
+    window_shift: int = 6,  # lfr_n
+):
+    opts = knf.FbankOptions()
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.window_type = "hamming"
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    online_fbank = knf.OnlineFbank(opts)
+    online_fbank.accept_waveform(sample_rate, (samples * 32768).tolist())
+    online_fbank.input_finished()
+
+    features = np.stack(
+        [online_fbank.get_frame(i) for i in range(online_fbank.num_frames_ready)]
+    )
+    assert features.data.contiguous is True
+    assert features.dtype == np.float32, features.dtype
+
+    T = (features.shape[0] - window_size) // window_shift + 1
+    features = np.lib.stride_tricks.as_strided(
+        features,
+        shape=(T, features.shape[1] * window_size),
+        strides=((window_shift * features.shape[1]) * 4, 4),
+    )
+
+    return np.copy(features)
+
+
+def main():
+    samples, sample_rate = load_audio("./test_wavs/zh.wav")
+    if sample_rate != 16000:
+        import librosa
+
+        samples = librosa.resample(samples, orig_sr=sample_rate, target_sr=16000)
+        sample_rate = 16000
+
+    model = OmModel()
+
+    features = compute_feat(
+        samples=samples,
+        sample_rate=sample_rate,
+    )
+    print("features.shape", features.shape)
+
+    language_auto = 0
+    language_zh = 3
+    language_en = 4
+    language_yue = 7
+    language_ya = 11
+    language_ko = 12
+    language_nospeech = 13
+
+    language = language_auto
+
+    with_itn = 14
+    without_itn = 15
+
+    text_norm = with_itn
+
+    prompt = np.array([language, 1, 2, text_norm], dtype=np.int32)
+    # language = np.array([language], dtype=np.int32)
+    # text_norm = np.array([text_norm], dtype=np.int32)
+
+    print("prompt", prompt.shape)
+
+    logits = model(
+        x=features[None],
+        prompt=prompt,
+        # language=language,
+        ##text_norm=text_norm,
+    )
+    print("logits.shape", logits.shape, type(logits))
+
+    idx = logits.argmax(axis=-1)
+    print(idx)
+    print(len(idx))
+    prev = -1
+    ids = []
+    for i in idx:
+        if i != prev:
+            ids.append(i)
+        prev = i
+    ids = [i for i in ids if i != 0]
+    print(ids)
+
+    tokens = load_tokens("./tokens.txt")
+    text = "".join([tokens[i] for i in ids])
+
+    text = text.replace("▁", " ")
+    print(text)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/sense-voice/ascend-npu/torch_model.py b/scripts/sense-voice/ascend-npu/torch_model.py
new file mode 120000
index 0000000000..2ed3ee1086
--- /dev/null
+++ b/scripts/sense-voice/ascend-npu/torch_model.py
@@ -0,0 +1 @@
+../rknn/torch_model.py
\ No newline at end of file