Add C++ runtime for silero_vad with RKNN (#2078)

2025-04-01 15:56:56 +08:00
parent 0703bc1b86
commit 8e51a97550
12 changed files with 867 additions and 16 deletions
--- a/c-api-examples/vad-whisper-c-api.c
+++ b/c-api-examples/vad-whisper-c-api.c
@@ -100,12 +100,11 @@ int32_t main() {

  while (!is_eof) {
    if (i + window_size < wave->num_samples) {
-        SherpaOnnxVoiceActivityDetectorAcceptWaveform(vad, wave->samples + i,
-            window_size);
-    }
-    else {
-        SherpaOnnxVoiceActivityDetectorFlush(vad);
-        is_eof = 1;
+      SherpaOnnxVoiceActivityDetectorAcceptWaveform(vad, wave->samples + i,
+                                                    window_size);
+    } else {
+      SherpaOnnxVoiceActivityDetectorFlush(vad);
+      is_eof = 1;
    }
    while (!SherpaOnnxVoiceActivityDetectorEmpty(vad)) {
      const SherpaOnnxSpeechSegment *segment =
--- a/scripts/gtcrn/show.py
+++ b/scripts/gtcrn/show.py
@@ -0,0 +1,75 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import onnxruntime
+import onnx
+
+"""
+[key: "model_type"
+value: "gtcrn"
+, key: "comment"
+value: "gtcrn_simple"
+, key: "version"
+value: "1"
+, key: "sample_rate"
+value: "16000"
+, key: "model_url"
+value: "https://github.com/Xiaobin-Rong/gtcrn/blob/main/stream/onnx_models/gtcrn_simple.onnx"
+, key: "maintainer"
+value: "k2-fsa"
+, key: "comment2"
+value: "Please see also https://github.com/Xiaobin-Rong/gtcrn"
+, key: "conv_cache_shape"
+value: "2,1,16,16,33"
+, key: "tra_cache_shape"
+value: "2,3,1,1,16"
+, key: "inter_cache_shape"
+value: "2,1,33,16"
+, key: "n_fft"
+value: "512"
+, key: "hop_length"
+value: "256"
+, key: "window_length"
+value: "512"
+, key: "window_type"
+value: "hann_sqrt"
+]
+"""
+
+"""
+NodeArg(name='mix', type='tensor(float)', shape=[1, 257, 1, 2])
+NodeArg(name='conv_cache', type='tensor(float)', shape=[2, 1, 16, 16, 33])
+NodeArg(name='tra_cache', type='tensor(float)', shape=[2, 3, 1, 1, 16])
+NodeArg(name='inter_cache', type='tensor(float)', shape=[2, 1, 33, 16])
+-----
+NodeArg(name='enh', type='tensor(float)', shape=[1, 257, 1, 2])
+NodeArg(name='conv_cache_out', type='tensor(float)', shape=[2, 1, 16, 16, 33])
+NodeArg(name='tra_cache_out', type='tensor(float)', shape=[2, 3, 1, 1, 16])
+NodeArg(name='inter_cache_out', type='tensor(float)', shape=[2, 1, 33, 16])
+"""
+
+
+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("./gtcrn_simple.onnx")
+
+
+if __name__ == "__main__":
+    main()
--- a/scripts/silero_vad/v4/export-onnx.py
+++ b/scripts/silero_vad/v4/export-onnx.py
@@ -5,15 +5,94 @@ import onnx
 import torch
 from onnxsim import simplify

+import torch
+from torch import Tensor
+
+
+def simple_pad(x: Tensor, pad: int) -> Tensor:
+    #  _0 = torch.slice(torch.slice(torch.slice(x), 1), 2, 1, torch.add(1, pad))
+    _0 = x[:, :, 1 : 1 + pad]
+
+    left_pad = torch.flip(_0, [-1])
+    #  _1 = torch.slice(torch.slice(torch.slice(x), 1), 2, torch.sub(-1, pad), -1)
+
+    _1 = x[:, :, (-1 - pad) : -1]
+
+    right_pad = torch.flip(_1, [-1])
+    _2 = torch.cat([left_pad, x, right_pad], 2)
+    return _2
+
+
+class MyModule(torch.nn.Module):
+    def __init__(self, m):
+        super().__init__()
+        self.m = m
+
+    def adaptive_normalization_forward(self, spect):
+        m = self.m._model.adaptive_normalization
+        _0 = simple_pad
+
+        # Note(fangjun): rknn uses fp16 by default, whose max value is 65504
+        # so we need to re-write the computation for spect0
+        #  spect0 = torch.log1p(torch.mul(spect, 1048576))
+        spect0 = torch.log1p(spect) + 13.86294
+
+        _1 = torch.eq(len(spect0.shape), 2)
+        if _1:
+            _2 = torch.unsqueeze(spect0, 0)
+            spect1 = _2
+        else:
+            spect1 = spect0
+        mean = torch.mean(spect1, [1], True)
+        to_pad = m.to_pad
+        mean0 = _0(
+            mean,
+            to_pad,
+        )
+        filter_ = m.filter_
+        mean1 = torch.conv1d(mean0, filter_)
+        mean_mean = torch.mean(mean1, [-1], True)
+        spect2 = torch.add(spect1, torch.neg(mean_mean))
+        return spect2
+
+    def forward(self, x: torch.Tensor, h: torch.Tensor, c: torch.Tensor):
+        m = self.m._model
+
+        feature_extractor = m.feature_extractor
+        x0 = (feature_extractor).forward(
+            x,
+        )
+        norm = self.adaptive_normalization_forward(x0)
+        x1 = torch.cat([x0, norm], 1)
+        first_layer = m.first_layer
+        x2 = (first_layer).forward(
+            x1,
+        )
+        encoder = m.encoder
+        x3 = (encoder).forward(
+            x2,
+        )
+        decoder = m.decoder
+        x4, h0, c0, = (decoder).forward(
+            x3,
+            h,
+            c,
+        )
+        _0 = torch.mean(torch.squeeze(x4, 1), [1])
+        out = torch.unsqueeze(_0, 1)
+        return (out, h0, c0)
+

@torch.no_grad()
 def main():
    m = torch.jit.load("./silero_vad.jit")
+    m = MyModule(m)
    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)
+    m = torch.jit.script(m)
    torch.onnx.export(
-        m._model,
+        m,
        (x, h, c),
        "m.onnx",
        input_names=["x", "h", "c"],
--- a/scripts/silero_vad/v4/show.py
+++ b/scripts/silero_vad/v4/show.py
@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-# Copyright    2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)

 import onnxruntime
 import onnx
--- a/scripts/silero_vad/v4/test-on-rk3588-board.py
+++ b/scripts/silero_vad/v4/test-on-rk3588-board.py
@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+# Please run this file on your rk3588 board
+
+try:
+    from rknnlite.api import RKNNLite
+except:
+    print("Please run this file on your board (linux + aarch64 + npu)")
+    print("You need to install rknn_toolkit_lite2")
+    print(
+        " from https://github.com/airockchip/rknn-toolkit2/tree/master/rknn-toolkit-lite2/packages"
+    )
+    print(
+        "https://github.com/airockchip/rknn-toolkit2/blob/v2.1.0/rknn-toolkit-lite2/packages/rknn_toolkit_lite2-2.1.0-cp310-cp310-linux_aarch64.whl"
+    )
+    print("is known to work")
+    raise
+
+import time
+from pathlib import Path
+from typing import Tuple
+
+import numpy as np
+import soundfile as sf
+
+
+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 init_model(filename, target_platform="rk3588"):
+    if not Path(filename).is_file():
+        exit(f"{filename} does not exist")
+
+    rknn_lite = RKNNLite(verbose=False)
+    ret = rknn_lite.load_rknn(path=filename)
+    if ret != 0:
+        exit(f"Load model {filename} failed!")
+
+    ret = rknn_lite.init_runtime(core_mask=RKNNLite.NPU_CORE_0)
+    if ret != 0:
+        exit(f"Failed to init rknn runtime for {filename}")
+    return rknn_lite
+
+
+class RKNNModel:
+    def __init__(self, model: str, target_platform="rk3588"):
+        self.model = init_model(model)
+
+    def release(self):
+        self.model.release()
+
+    def __call__(self, x: np.ndarray, h: np.ndarray, c: np.ndarray):
+        """
+        Args:
+          x: (1, 512), np.float32
+          h: (2, 1, 64), np.float32
+          c: (2, 1, 64), np.float32
+        Returns:
+          prob:
+          next_h:
+          next_c
+        """
+        out, next_h, next_c = self.model.inference(inputs=[x, h, c])
+        return out.item(), next_h, next_c
+
+
+def main():
+    model = RKNNModel(model="./m.rknn")
+    for i in range(1):
+        test(model)
+
+
+def test(model):
+    print("started")
+    start = time.time()
+    samples, sample_rate = load_audio("./lei-jun-test.wav")
+    assert sample_rate == 16000, sample_rate
+
+    window_size = 512
+
+    h = np.zeros((2, 1, 64), dtype=np.float32)
+    c = np.zeros((2, 1, 64), dtype=np.float32)
+
+    threshold = 0.5
+    num_windows = samples.shape[0] // window_size
+    out = []
+    for i in range(num_windows):
+        print(i, num_windows)
+        this_samples = samples[i * window_size : (i + 1) * window_size]
+        prob, h, c = model(this_samples[None], h, c)
+        out.append(prob > threshold)
+
+    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("Empty for ./lei-jun-test.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()
--- a/scripts/silero_vad/v4/test-onnx.py
+++ b/scripts/silero_vad/v4/test-onnx.py
@@ -97,10 +97,13 @@ def main():
    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
--- a/sherpa-onnx/csrc/CMakeLists.txt
+++ b/sherpa-onnx/csrc/CMakeLists.txt
@@ -159,6 +159,7 @@ if(SHERPA_ONNX_ENABLE_RKNN)
    ./rknn/online-transducer-modified-beam-search-decoder-rknn.cc
    ./rknn/online-zipformer-ctc-model-rknn.cc
    ./rknn/online-zipformer-transducer-model-rknn.cc
+    ./rknn/silero-vad-model-rknn.cc
    ./rknn/utils.cc
  )

@@ -468,6 +469,7 @@ if(SHERPA_ONNX_ENABLE_PORTAUDIO AND SHERPA_ONNX_ENABLE_BINARY)
    microphone.cc
  )

+
  add_executable(sherpa-onnx-microphone-offline
    sherpa-onnx-microphone-offline.cc
    microphone.cc
@@ -498,11 +500,11 @@ if(SHERPA_ONNX_ENABLE_PORTAUDIO AND SHERPA_ONNX_ENABLE_BINARY)
  )

  set(exes
-    sherpa-onnx-microphone
    sherpa-onnx-keyword-spotter-microphone
+    sherpa-onnx-microphone
    sherpa-onnx-microphone-offline
-    sherpa-onnx-microphone-offline-speaker-identification
    sherpa-onnx-microphone-offline-audio-tagging
+    sherpa-onnx-microphone-offline-speaker-identification
    sherpa-onnx-vad-microphone
    sherpa-onnx-vad-microphone-offline-asr
    sherpa-onnx-vad-with-offline-asr
--- a/sherpa-onnx/csrc/rknn/silero-vad-model-rknn.cc
+++ b/sherpa-onnx/csrc/rknn/silero-vad-model-rknn.cc
@@ -0,0 +1,470 @@
+// sherpa-onnx/csrc/rknn/silero-vad-model-rknn.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/rknn/silero-vad-model-rknn.h"
+
+#include <string>
+#include <utility>
+#include <vector>
+
+#if __ANDROID_API__ >= 9
+#include "android/asset_manager.h"
+#include "android/asset_manager_jni.h"
+#endif
+
+#if __OHOS__
+#include "rawfile/raw_file_manager.h"
+#endif
+
+#include "sherpa-onnx/csrc/file-utils.h"
+#include "sherpa-onnx/csrc/macros.h"
+#include "sherpa-onnx/csrc/rknn/macros.h"
+#include "sherpa-onnx/csrc/rknn/utils.h"
+#include "sherpa-onnx/csrc/text-utils.h"
+
+namespace sherpa_onnx {
+
+class SileroVadModelRknn::Impl {
+ public:
+  ~Impl() {
+    auto ret = rknn_destroy(ctx_);
+    if (ret != RKNN_SUCC) {
+      SHERPA_ONNX_LOGE("Failed to destroy the context");
+    }
+  }
+
+  explicit Impl(const VadModelConfig &config)
+      : config_(config), sample_rate_(config.sample_rate) {
+    auto buf = ReadFile(config.silero_vad.model);
+    Init(buf.data(), buf.size());
+
+    if (sample_rate_ != 16000) {
+      SHERPA_ONNX_LOGE("Expected sample rate 16000. Given: %d",
+                       config.sample_rate);
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    min_silence_samples_ =
+        sample_rate_ * config_.silero_vad.min_silence_duration;
+
+    min_speech_samples_ = sample_rate_ * config_.silero_vad.min_speech_duration;
+  }
+
+  template <typename Manager>
+  Impl(Manager *mgr, const VadModelConfig &config)
+      : config_(config), sample_rate_(config.sample_rate) {
+    auto buf = ReadFile(mgr, config.silero_vad.model);
+    Init(buf.data(), buf.size());
+
+    if (sample_rate_ != 16000) {
+      SHERPA_ONNX_LOGE("Expected sample rate 16000. Given: %d",
+                       config.sample_rate);
+      exit(-1);
+    }
+
+    min_silence_samples_ =
+        sample_rate_ * config_.silero_vad.min_silence_duration;
+
+    min_speech_samples_ = sample_rate_ * config_.silero_vad.min_speech_duration;
+  }
+
+  void Reset() {
+    for (auto &s : states_) {
+      std::fill(s.begin(), s.end(), 0);
+    }
+
+    triggered_ = false;
+    current_sample_ = 0;
+    temp_start_ = 0;
+    temp_end_ = 0;
+  }
+
+  bool IsSpeech(const float *samples, int32_t n) {
+    if (n != WindowSize()) {
+      SHERPA_ONNX_LOGE("n: %d != window_size: %d", n, WindowSize());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    float prob = Run(samples, n);
+
+    float threshold = config_.silero_vad.threshold;
+
+    current_sample_ += config_.silero_vad.window_size;
+
+    if (prob > threshold && temp_end_ != 0) {
+      temp_end_ = 0;
+    }
+
+    if (prob > threshold && temp_start_ == 0) {
+      // start speaking, but we require that it must satisfy
+      // min_speech_duration
+      temp_start_ = current_sample_;
+      return false;
+    }
+
+    if (prob > threshold && temp_start_ != 0 && !triggered_) {
+      if (current_sample_ - temp_start_ < min_speech_samples_) {
+        return false;
+      }
+
+      triggered_ = true;
+
+      return true;
+    }
+
+    if ((prob < threshold) && !triggered_) {
+      // silence
+      temp_start_ = 0;
+      temp_end_ = 0;
+      return false;
+    }
+
+    if ((prob > threshold - 0.15) && triggered_) {
+      // speaking
+      return true;
+    }
+
+    if ((prob > threshold) && !triggered_) {
+      // start speaking
+      triggered_ = true;
+
+      return true;
+    }
+
+    if ((prob < threshold) && triggered_) {
+      // stop to speak
+      if (temp_end_ == 0) {
+        temp_end_ = current_sample_;
+      }
+
+      if (current_sample_ - temp_end_ < min_silence_samples_) {
+        // continue speaking
+        return true;
+      }
+      // stopped speaking
+      temp_start_ = 0;
+      temp_end_ = 0;
+      triggered_ = false;
+      return false;
+    }
+
+    return false;
+  }
+
+  int32_t WindowShift() const { return config_.silero_vad.window_size; }
+
+  int32_t WindowSize() const {
+    return config_.silero_vad.window_size + window_overlap_;
+  }
+
+  int32_t MinSilenceDurationSamples() const { return min_silence_samples_; }
+
+  int32_t MinSpeechDurationSamples() const { return min_speech_samples_; }
+
+  void SetMinSilenceDuration(float s) {
+    min_silence_samples_ = sample_rate_ * s;
+  }
+
+  void SetThreshold(float threshold) {
+    config_.silero_vad.threshold = threshold;
+  }
+
+ private:
+  void Init(void *model_data, size_t model_data_length) {
+    auto ret = rknn_init(&ctx_, model_data, model_data_length, 0, nullptr);
+    SHERPA_ONNX_RKNN_CHECK(ret, "Failed to init silero vad model '%s'",
+                           config_.silero_vad.model.c_str());
+
+    if (config_.debug) {
+      rknn_sdk_version v;
+      ret = rknn_query(ctx_, RKNN_QUERY_SDK_VERSION, &v, sizeof(v));
+      SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get rknn sdk version");
+
+      SHERPA_ONNX_LOGE("sdk api version: %s, driver version: %s", v.api_version,
+                       v.drv_version);
+    }
+
+    rknn_input_output_num io_num;
+    ret = rknn_query(ctx_, RKNN_QUERY_IN_OUT_NUM, &io_num, sizeof(io_num));
+    SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get I/O information for the model");
+
+    if (config_.debug) {
+      SHERPA_ONNX_LOGE("model: %d inputs, %d outputs",
+                       static_cast<int32_t>(io_num.n_input),
+                       static_cast<int32_t>(io_num.n_output));
+    }
+
+    input_attrs_.resize(io_num.n_input);
+    output_attrs_.resize(io_num.n_output);
+
+    int32_t i = 0;
+    for (auto &attr : input_attrs_) {
+      memset(&attr, 0, sizeof(attr));
+      attr.index = i;
+      ret = rknn_query(ctx_, RKNN_QUERY_INPUT_ATTR, &attr, sizeof(attr));
+      SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for model input %d", i);
+      i += 1;
+    }
+
+    if (config_.debug) {
+      std::ostringstream os;
+      std::string sep;
+      for (auto &attr : input_attrs_) {
+        os << sep << ToString(attr);
+        sep = "\n";
+      }
+      SHERPA_ONNX_LOGE("\n----------Model inputs info----------\n%s",
+                       os.str().c_str());
+    }
+
+    i = 0;
+    for (auto &attr : output_attrs_) {
+      memset(&attr, 0, sizeof(attr));
+      attr.index = i;
+      ret = rknn_query(ctx_, RKNN_QUERY_OUTPUT_ATTR, &attr, sizeof(attr));
+      SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for model output %d", i);
+      i += 1;
+    }
+
+    if (config_.debug) {
+      std::ostringstream os;
+      std::string sep;
+      for (auto &attr : output_attrs_) {
+        os << sep << ToString(attr);
+        sep = "\n";
+      }
+      SHERPA_ONNX_LOGE("\n----------Model outputs info----------\n%s",
+                       os.str().c_str());
+    }
+
+    rknn_custom_string custom_string;
+    ret = rknn_query(ctx_, RKNN_QUERY_CUSTOM_STRING, &custom_string,
+                     sizeof(custom_string));
+    SHERPA_ONNX_RKNN_CHECK(ret, "Failed to read custom string from the model");
+    if (config_.debug) {
+      SHERPA_ONNX_LOGE("customs string: %s", custom_string.string);
+    }
+    auto meta = Parse(custom_string);
+
+    if (config_.silero_vad.window_size != 512) {
+      SHERPA_ONNX_LOGE("we require window_size to be 512. Given: %d",
+                       config_.silero_vad.window_size);
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (config_.debug) {
+      for (const auto &p : meta) {
+        SHERPA_ONNX_LOGE("%s: %s", p.first.c_str(), p.second.c_str());
+      }
+    }
+
+    if (meta.count("model_type") == 0) {
+      SHERPA_ONNX_LOGE("No model type found in '%s'",
+                       config_.silero_vad.model.c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (meta.at("model_type") != "silero-vad-v4") {
+      SHERPA_ONNX_LOGE("Expect model type silero-vad-v4 in '%s', given: '%s'",
+                       config_.silero_vad.model.c_str(),
+                       meta.at("model_type").c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (meta.count("sample_rate") == 0) {
+      SHERPA_ONNX_LOGE("No sample_rate found in '%s'",
+                       config_.silero_vad.model.c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (meta.at("sample_rate") != "16000") {
+      SHERPA_ONNX_LOGE("Expect sample rate 16000 in '%s', given: '%s'",
+                       config_.silero_vad.model.c_str(),
+                       meta.at("sample_rate").c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (meta.count("version") == 0) {
+      SHERPA_ONNX_LOGE("No version found in '%s'",
+                       config_.silero_vad.model.c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (meta.at("version") != "4") {
+      SHERPA_ONNX_LOGE("Expect version 4 in '%s', given: '%s'",
+                       config_.silero_vad.model.c_str(),
+                       meta.at("version").c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (meta.count("h_shape") == 0) {
+      SHERPA_ONNX_LOGE("No h_shape found in '%s'",
+                       config_.silero_vad.model.c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    if (meta.count("c_shape") == 0) {
+      SHERPA_ONNX_LOGE("No c_shape found in '%s'",
+                       config_.silero_vad.model.c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    std::vector<int64_t> h_shape;
+    std::vector<int64_t> c_shape;
+
+    SplitStringToIntegers(meta.at("h_shape"), ",", false, &h_shape);
+    SplitStringToIntegers(meta.at("c_shape"), ",", false, &c_shape);
+    if (h_shape.size() != 3 || c_shape.size() != 3) {
+      SHERPA_ONNX_LOGE("Incorrect shape for h (%d) or c (%d)",
+                       static_cast<int32_t>(h_shape.size()),
+                       static_cast<int32_t>(c_shape.size()));
+      SHERPA_ONNX_EXIT(-1);
+    }
+
+    states_.resize(2);
+    states_[0].resize(h_shape[0] * h_shape[1] * h_shape[2]);
+    states_[1].resize(c_shape[0] * c_shape[1] * c_shape[2]);
+
+    Reset();
+  }
+
+  float Run(const float *samples, int32_t n) {
+    std::vector<rknn_input> inputs(input_attrs_.size());
+
+    for (int32_t i = 0; i < static_cast<int32_t>(inputs.size()); ++i) {
+      auto &input = inputs[i];
+      auto &attr = input_attrs_[i];
+      input.index = attr.index;
+
+      if (attr.type == RKNN_TENSOR_FLOAT16) {
+        input.type = RKNN_TENSOR_FLOAT32;
+      } else if (attr.type == RKNN_TENSOR_INT64) {
+        input.type = RKNN_TENSOR_INT64;
+      } else {
+        SHERPA_ONNX_LOGE("Unsupported tensor type %d, %s", attr.type,
+                         get_type_string(attr.type));
+        SHERPA_ONNX_EXIT(-1);
+      }
+
+      input.fmt = attr.fmt;
+      if (i == 0) {
+        input.buf = reinterpret_cast<void *>(const_cast<float *>(samples));
+        input.size = n * sizeof(float);
+      } else {
+        input.buf = reinterpret_cast<void *>(states_[i - 1].data());
+        input.size = states_[i - 1].size() * sizeof(float);
+      }
+    }
+
+    std::vector<float> out(output_attrs_[0].n_elems);
+
+    auto &next_states = states_;
+
+    std::vector<rknn_output> outputs(output_attrs_.size());
+
+    for (int32_t i = 0; i < outputs.size(); ++i) {
+      auto &output = outputs[i];
+      auto &attr = output_attrs_[i];
+      output.index = attr.index;
+      output.is_prealloc = 1;
+
+      if (attr.type == RKNN_TENSOR_FLOAT16) {
+        output.want_float = 1;
+      } else if (attr.type == RKNN_TENSOR_INT64) {
+        output.want_float = 0;
+      } else {
+        SHERPA_ONNX_LOGE("Unsupported tensor type %d, %s", attr.type,
+                         get_type_string(attr.type));
+        SHERPA_ONNX_EXIT(-1);
+      }
+
+      if (i == 0) {
+        output.size = out.size() * sizeof(float);
+        output.buf = reinterpret_cast<void *>(out.data());
+      } else {
+        output.size = next_states[i - 1].size() * sizeof(float);
+        output.buf = reinterpret_cast<void *>(next_states[i - 1].data());
+      }
+    }
+
+    auto ret = rknn_inputs_set(ctx_, inputs.size(), inputs.data());
+    SHERPA_ONNX_RKNN_CHECK(ret, "Failed to set inputs");
+
+    ret = rknn_run(ctx_, nullptr);
+    SHERPA_ONNX_RKNN_CHECK(ret, "Failed to run the model");
+
+    ret = rknn_outputs_get(ctx_, outputs.size(), outputs.data(), nullptr);
+    SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get model output");
+
+    return out[0];
+  }
+
+ private:
+  VadModelConfig config_;
+  rknn_context ctx_ = 0;
+
+  std::vector<rknn_tensor_attr> input_attrs_;
+  std::vector<rknn_tensor_attr> output_attrs_;
+
+  std::vector<std::vector<float>> states_;
+
+  int64_t sample_rate_;
+  int32_t min_silence_samples_;
+  int32_t min_speech_samples_;
+
+  bool triggered_ = false;
+  int32_t current_sample_ = 0;
+  int32_t temp_start_ = 0;
+  int32_t temp_end_ = 0;
+
+  int32_t window_overlap_ = 0;
+};
+
+SileroVadModelRknn::SileroVadModelRknn(const VadModelConfig &config)
+    : impl_(std::make_unique<Impl>(config)) {}
+
+template <typename Manager>
+SileroVadModelRknn::SileroVadModelRknn(Manager *mgr,
+                                       const VadModelConfig &config)
+    : impl_(std::make_unique<Impl>(mgr, config)) {}
+
+SileroVadModelRknn::~SileroVadModelRknn() = default;
+
+void SileroVadModelRknn::Reset() { return impl_->Reset(); }
+
+bool SileroVadModelRknn::IsSpeech(const float *samples, int32_t n) {
+  return impl_->IsSpeech(samples, n);
+}
+
+int32_t SileroVadModelRknn::WindowSize() const { return impl_->WindowSize(); }
+
+int32_t SileroVadModelRknn::WindowShift() const { return impl_->WindowShift(); }
+
+int32_t SileroVadModelRknn::MinSilenceDurationSamples() const {
+  return impl_->MinSilenceDurationSamples();
+}
+
+int32_t SileroVadModelRknn::MinSpeechDurationSamples() const {
+  return impl_->MinSpeechDurationSamples();
+}
+
+void SileroVadModelRknn::SetMinSilenceDuration(float s) {
+  impl_->SetMinSilenceDuration(s);
+}
+
+void SileroVadModelRknn::SetThreshold(float threshold) {
+  impl_->SetThreshold(threshold);
+}
+
+#if __ANDROID_API__ >= 9
+template SileroVadModelRknn::SileroVadModelRknn(AAssetManager *mgr,
+                                                const VadModelConfig &config);
+#endif
+
+#if __OHOS__
+template SileroVadModelRknn::SileroVadModelRknn(NativeResourceManager *mgr,
+                                                const VadModelConfig &config);
+#endif
+
+}  // namespace sherpa_onnx
--- a/sherpa-onnx/csrc/rknn/silero-vad-model-rknn.h
+++ b/sherpa-onnx/csrc/rknn/silero-vad-model-rknn.h
@@ -0,0 +1,53 @@
+// sherpa-onnx/csrc/rknn/silero-vad-model-rknn.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+#ifndef SHERPA_ONNX_CSRC_RKNN_SILERO_VAD_MODEL_RKNN_H_
+#define SHERPA_ONNX_CSRC_RKNN_SILERO_VAD_MODEL_RKNN_H_
+
+#include "rknn_api.h"  // NOLINT
+#include "sherpa-onnx/csrc/online-model-config.h"
+#include "sherpa-onnx/csrc/vad-model.h"
+
+namespace sherpa_onnx {
+
+class SileroVadModelRknn : public VadModel {
+ public:
+  explicit SileroVadModelRknn(const VadModelConfig &config);
+
+  template <typename Manager>
+  SileroVadModelRknn(Manager *mgr, const VadModelConfig &config);
+
+  ~SileroVadModelRknn() override;
+
+  // reset the internal model states
+  void Reset() override;
+
+  /**
+   * @param samples Pointer to a 1-d array containing audio samples.
+   *                Each sample should be normalized to the range [-1, 1].
+   * @param n Number of samples.
+   *
+   * @return Return true if speech is detected. Return false otherwise.
+   */
+  bool IsSpeech(const float *samples, int32_t n) override;
+
+  // For silero vad V4, it is WindowShift().
+  int32_t WindowSize() const override;
+
+  // 512
+  int32_t WindowShift() const override;
+
+  int32_t MinSilenceDurationSamples() const override;
+  int32_t MinSpeechDurationSamples() const override;
+
+  void SetMinSilenceDuration(float s) override;
+  void SetThreshold(float threshold) override;
+
+ private:
+  class Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_RKNN_SILERO_VAD_MODEL_RKNN_H_
--- a/sherpa-onnx/csrc/sherpa-onnx-vad-alsa-offline-asr.cc
+++ b/sherpa-onnx/csrc/sherpa-onnx-vad-alsa-offline-asr.cc
@@ -129,15 +129,13 @@ as the device_name.
    exit(-1);
  }

-  int32_t chunk = 0.1 * alsa.GetActualSampleRate();
-
  fprintf(stderr, "Started. Please speak\n");

  int32_t window_size = vad_config.silero_vad.window_size;
  int32_t index = 0;

  while (!stop) {
-    const std::vector<float> &samples = alsa.Read(chunk);
+    const std::vector<float> &samples = alsa.Read(window_size);
    vad->AcceptWaveform(samples.data(), samples.size());

    while (!vad->Empty()) {
--- a/sherpa-onnx/csrc/vad-model-config.cc
+++ b/sherpa-onnx/csrc/vad-model-config.cc
@@ -7,6 +7,9 @@
 #include <sstream>
 #include <string>

+#include "sherpa-onnx/csrc/macros.h"
+#include "sherpa-onnx/csrc/text-utils.h"
+
 namespace sherpa_onnx {

 void VadModelConfig::Register(ParseOptions *po) {
@@ -26,7 +29,27 @@ void VadModelConfig::Register(ParseOptions *po) {
               "true to display debug information when loading vad models");
 }

-bool VadModelConfig::Validate() const { return silero_vad.Validate(); }
+bool VadModelConfig::Validate() const {
+  if (provider != "rknn") {
+    if (!silero_vad.model.empty() && EndsWith(silero_vad.model, ".rknn")) {
+      SHERPA_ONNX_LOGE(
+          "--provider is %s, which is not rknn, but you pass an rknn model "
+          "'%s'",
+          provider.c_str(), silero_vad.model.c_str());
+      return false;
+    }
+  }
+
+  if (provider == "rknn") {
+    if (!silero_vad.model.empty() && EndsWith(silero_vad.model, ".onnx")) {
+      SHERPA_ONNX_LOGE("--provider is rknn, but you pass an onnx model '%s'",
+                       silero_vad.model.c_str());
+      return false;
+    }
+  }
+
+  return silero_vad.Validate();
+}

 std::string VadModelConfig::ToString() const {
  std::ostringstream os;
--- a/sherpa-onnx/csrc/vad-model.cc
+++ b/sherpa-onnx/csrc/vad-model.cc
@@ -13,19 +13,27 @@
 #include "rawfile/raw_file_manager.h"
 #endif

+#if SHERPA_ONNX_ENABLE_RKNN
+#include "sherpa-onnx/csrc/rknn/silero-vad-model-rknn.h"
+#endif
+
 #include "sherpa-onnx/csrc/silero-vad-model.h"

 namespace sherpa_onnx {

 std::unique_ptr<VadModel> VadModel::Create(const VadModelConfig &config) {
-  // TODO(fangjun): Support other VAD models.
+  if (config.provider == "rknn") {
+    return std::make_unique<SileroVadModelRknn>(config);
+  }
  return std::make_unique<SileroVadModel>(config);
 }

 template <typename Manager>
 std::unique_ptr<VadModel> VadModel::Create(Manager *mgr,
                                           const VadModelConfig &config) {
-  // TODO(fangjun): Support other VAD models.
+  if (config.provider == "rknn") {
+    return std::make_unique<SileroVadModelRknn>(mgr, config);
+  }
  return std::make_unique<SileroVadModel>(mgr, config);
 }