Export silero_vad v4 to RKNN (#2067)

+name: export-silero-vad-to-rknn
+
+on:
+  workflow_dispatch:
+
+concurrency:
+  group: export-silero-vad-to-rknn-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  export-silero-vad-to-rknn:
+    if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
+    name: export silero-vad to rknn
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-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 Python dependencies
+        shell: bash
+        run: |
+          python3 -m pip install --upgrade \
+            pip \
+            "numpy<2" \
+            torch==2.0.0+cpu -f https://download.pytorch.org/whl/torch \
+            onnx \
+            onnxruntime==1.17.1 \
+            librosa \
+            soundfile \
+            onnxsim
+
+          curl -SL -O https://huggingface.co/csukuangfj/rknn-toolkit2/resolve/main/rknn_toolkit2-2.1.0%2B708089d1-cp310-cp310-linux_x86_64.whl
+          pip install ./*.whl "numpy<=1.26.4"
+
+      - name: Run
+        shell: bash
+        run: |
+          cd scripts/silero_vad/v4
+          curl -SL -O https://github.com/snakers4/silero-vad/raw/refs/tags/v4.0/files/silero_vad.jit
+          ./export-onnx.py
+          ./show.py
+
+          ls -lh m.onnx
+
+          curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/lei-jun-test.wav
+          ./test-onnx.py  --model ./m.onnx --wav ./lei-jun-test.wav
+
+          for platform in rk3588 rk3576 rk3568 rk3566 rk3562; do
+          echo "Platform: $platform"
+            ./export-rknn.py --in-model ./m.onnx --out-model silero-vad-v4-$platform.rknn  --target-platform $platform
+            ls -lh silero-vad-v4-$platform.rknn
+          done
+
+      - name: Collect files
+        shell: bash
+        run: |
+          cd scripts/silero_vad/v4
+          ls -lh
+          mv *.rknn ../../..
+
+      - name: Release
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.rknn
+          overwrite: true
+          repo_name: k2-fsa/sherpa-onnx
+          repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
+          tag: asr-models
+
+      - name: Upload model to huggingface
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        uses: nick-fields/retry@v3
+        with:
+          max_attempts: 20
+          timeout_seconds: 200
+          shell: bash
+          command: |
+            git config --global user.email "csukuangfj@gmail.com"
+            git config --global user.name "Fangjun Kuang"
+
+            rm -rf huggingface
+            export GIT_LFS_SKIP_SMUDGE=1
+
+            git clone https://huggingface.co/csukuangfj/sherpa-onnx-rknn-models huggingface
+            cd huggingface
+
+            git fetch
+            git pull
+            git lfs track "*.rknn"
+            git merge -m "merge remote" --ff origin main
+            dst=vad
+            mkdir -p $dst
+            cp ../*.rknn $dst/ || true
+
+            ls -lh $dst
+            git add .
+            git status
+            git commit -m "update models"
+            git status
+
+            git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-rknn-models main || true
+            rm -rf huggingface

@@ -136,6 +136,7 @@ kokoro-multi-lang-v1_0
 sherpa-onnx-fire-red-asr-large-zh_en-2025-02-16
 cmake-build-debug
 README-DEV.txt
-
+*.rknn
+*.jit
 ##clion
 .idea

+# Introduction
+
+This folder contains script for exporting
+[silero_vad v4](https://github.com/snakers4/silero-vad/tree/v4.0)
+to rknn.
+
+# Steps to run
+
+## 1. Download a jit model
+You can download it from <https://github.com/snakers4/silero-vad/blob/v4.0/files/silero_vad.jit>
+
+```bash
+wget https://github.com/snakers4/silero-vad/raw/refs/tags/v4.0/files/silero_vad.jit
+```
+
+```bash
+ls -lh silero_vad.jit
+-rw-r--r-- 1 kuangfangjun root 1.4M Mar 30 11:04 silero_vad.jit
+```
+
+## 2. Export it to onnx
+```bash
+./export-onnx.py
+```
+
+It will generate a file `./m.onnx`
+
+```bash
+ ls -lh m.onnx
+-rw-r--r-- 1 kuangfangjun root 627K Mar 30 11:13 m.onnx
+```
+
+## 3. Test the onnx model
+
+```bash
+wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/lei-jun-test.wav
+./test-onnx.py  --model ./m.onnx --wav ./lei-jun-test.wav
+```
+
+## 4. Convert the onnx model to RKNN format
+
+We assume you have installed rknn toolkit 2.1
+```bash
+./export-rknn.py --in-model ./m.onnx --out-model m.rknn  --target-platform rk3588
+```
+
+It will generate a file `./m.rknn`
+
+```bash
+ls -lh m.rknn
+-rw-r--r-- 1 kuangfangjun root 2.2M Mar 30 11:19 m.rknn
+```

+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import onnx
+import torch
+from onnxsim import simplify
+
+
+@torch.no_grad()
+def main():
+    m = torch.jit.load("./silero_vad.jit")
+    x = torch.rand((1, 512), dtype=torch.float32)
+    h = torch.rand((2, 1, 64), dtype=torch.float32)
+    c = torch.rand((2, 1, 64), dtype=torch.float32)
+    torch.onnx.export(
+        m._model,
+        (x, h, c),
+        "m.onnx",
+        input_names=["x", "h", "c"],
+        output_names=["prob", "next_h", "next_c"],
+    )
+
+    print("simplifying ...")
+    model = onnx.load("m.onnx")
+
+    meta_data = {
+        "model_type": "silero-vad-v4",
+        "sample_rate": 16000,
+        "version": 4,
+        "h_shape": "2,1,64",
+        "c_shape": "2,1,64",
+    }
+
+    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)
+    print("--------------------")
+    print(model.metadata_props)
+
+    model_simp, check = simplify(model)
+    onnx.save(model_simp, "m.onnx")
+
+
+if __name__ == "__main__":
+    main()

+#!/usr/bin/env python3
+# Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import argparse
+import logging
+from pathlib import Path
+
+from rknn.api import RKNN
+
+logging.basicConfig(level=logging.WARNING)
+
+g_platforms = [
+    #  "rv1103",
+    #  "rv1103b",
+    #  "rv1106",
+    #  "rk2118",
+    "rk3562",
+    "rk3566",
+    "rk3568",
+    "rk3576",
+    "rk3588",
+]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--target-platform",
+        type=str,
+        required=True,
+        help=f"Supported values are: {','.join(g_platforms)}",
+    )
+
+    parser.add_argument(
+        "--in-model",
+        type=str,
+        required=True,
+        help="Path to the input onnx model",
+    )
+
+    parser.add_argument(
+        "--out-model",
+        type=str,
+        required=True,
+        help="Path to the output rknn model",
+    )
+
+    return parser
+
+
+def get_meta_data(model: str):
+    import onnxruntime
+
+    session_opts = onnxruntime.SessionOptions()
+    session_opts.inter_op_num_threads = 1
+    session_opts.intra_op_num_threads = 1
+
+    m = onnxruntime.InferenceSession(
+        model,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    for i in m.get_inputs():
+        print(i)
+
+    print("-----")
+
+    for i in m.get_outputs():
+        print(i)
+    print()
+
+    meta = m.get_modelmeta().custom_metadata_map
+    s = ""
+    sep = ""
+    for key, value in meta.items():
+        s = s + sep + f"{key}={value}"
+        sep = ";"
+    assert len(s) < 1024
+
+    return s
+
+
+def export_rknn(rknn, filename):
+    ret = rknn.export_rknn(filename)
+    if ret != 0:
+        exit("Export rknn model to {filename} failed!")
+
+
+def init_model(filename: str, target_platform: str, custom_string=None):
+    rknn = RKNN(verbose=False)
+
+    rknn.config(
+        optimization_level=0,
+        target_platform=target_platform,
+        custom_string=custom_string,
+    )
+    if not Path(filename).is_file():
+        exit(f"{filename} does not exist")
+
+    ret = rknn.load_onnx(model=filename)
+    if ret != 0:
+        exit(f"Load model {filename} failed!")
+
+    ret = rknn.build(do_quantization=False)
+    if ret != 0:
+        exit("Build model {filename} failed!")
+
+    return rknn
+
+
+class RKNNModel:
+    def __init__(
+        self,
+        model: str,
+        target_platform: str,
+    ):
+        meta = get_meta_data(model)
+        print(meta)
+
+        self.model = init_model(
+            model,
+            target_platform=target_platform,
+            custom_string=meta,
+        )
+
+    def export_rknn(self, model):
+        export_rknn(self.model, model)
+
+    def release(self):
+        self.model.release()
+
+
+def main():
+    args = get_parser().parse_args()
+    print(vars(args))
+
+    model = RKNNModel(
+        model=args.in_model,
+        target_platform=args.target_platform,
+    )
+
+    model.export_rknn(
+        model=args.out_model,
+    )
+
+    model.release()
+
+
+if __name__ == "__main__":
+    main()

+#!/usr/bin/env python3
+# Copyright    2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import onnxruntime
+import onnx
+
+"""
+[key: "model_type"
+value: "silero-vad-v4"
+, key: "sample_rate"
+value: "16000"
+, key: "version"
+value: "4"
+, key: "h_shape"
+value: "2,1,64"
+, key: "c_shape"
+value: "2,1,64"
+]
+NodeArg(name='x', type='tensor(float)', shape=[1, 512])
+NodeArg(name='h', type='tensor(float)', shape=[2, 1, 64])
+NodeArg(name='c', type='tensor(float)', shape=[2, 1, 64])
+-----
+NodeArg(name='prob', type='tensor(float)', shape=[1, 1])
+NodeArg(name='next_h', type='tensor(float)', shape=[2, 1, 64])
+NodeArg(name='next_c', type='tensor(float)', shape=[2, 1, 64])
+"""
+
+
+def show(filename):
+    model = onnx.load(filename)
+    print(model.metadata_props)
+
+    session_opts = onnxruntime.SessionOptions()
+    session_opts.log_severity_level = 3
+    sess = onnxruntime.InferenceSession(
+        filename, session_opts, providers=["CPUExecutionProvider"]
+    )
+    for i in sess.get_inputs():
+        print(i)
+
+    print("-----")
+
+    for i in sess.get_outputs():
+        print(i)
+
+
+def main():
+    show("./m.onnx")
+
+
+if __name__ == "__main__":
+    main()

+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+import onnxruntime as ort
+import argparse
+import soundfile as sf
+from typing import Tuple
+import numpy as np
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Path to the onnx model",
+    )
+
+    parser.add_argument(
+        "--wav",
+        type=str,
+        required=True,
+        help="Path to the input wav",
+    )
+    return parser.parse_args()
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+        self.model = ort.InferenceSession(
+            model,
+            sess_options=session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def get_init_states(self):
+        h = np.zeros((2, 1, 64), dtype=np.float32)
+        c = np.zeros((2, 1, 64), dtype=np.float32)
+        return h, c
+
+    def __call__(self, x, h, c):
+        """
+        Args:
+          x: (1, 512)
+          h: (2, 1, 64)
+          c: (2, 1, 64)
+        Returns:
+          prob: (1, 1)
+          next_h: (2, 1, 64)
+          next_c: (2, 1, 64)
+        """
+        x = x[None]
+        out, next_h, next_c = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+                self.model.get_outputs()[2].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x,
+                self.model.get_inputs()[1].name: h,
+                self.model.get_inputs()[2].name: c,
+            },
+        )
+        return out, next_h, next_c
+
+
+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 main():
+    args = get_args()
+
+    samples, sample_rate = load_audio(args.wav)
+    if sample_rate != 16000:
+        import librosa
+
+        samples = librosa.resample(samples, orig_sr=sample_rate, target_sr=16000)
+        sample_rate = 16000
+
+    model = OnnxModel(args.model)
+    probs = []
+    h, c = model.get_init_states()
+    window_size = 512
+    num_windows = samples.shape[0] // window_size
+    for i in range(num_windows):
+        start = i * window_size
+        end = start + window_size
+        p, h, c = model(samples[start:end], h, c)
+        probs.append(p[0].item())
+
+    threshold = 0.5
+    out = np.array(probs) > threshold
+    out = out.tolist()
+    min_speech_duration = 0.25 * sample_rate / window_size
+    min_silence_duration = 0.25 * sample_rate / window_size
+
+    result = []
+    last = -1
+    for k, f in enumerate(out):
+        if f >= threshold:
+            if last == -1:
+                last = k
+        elif last != -1:
+            if k - last > min_speech_duration:
+                result.append((last, k))
+            last = -1
+
+    if last != -1 and k - last > min_speech_duration:
+        result.append((last, k))
+
+    if not result:
+        print(f"Empty for {args.wav}")
+        return
+
+    print(result)
+
+    final = [result[0]]
+    for r in result[1:]:
+        f = final[-1]
+        if r[0] - f[1] < min_silence_duration:
+            final[-1] = (f[0], r[1])
+        else:
+            final.append(r)
+
+    for f in final:
+        start = f[0] * window_size / sample_rate
+        end = f[1] * window_size / sample_rate
+        print("{:.3f} -- {:.3f}".format(start, end))
+
+
+if __name__ == "__main__":
+    main()