diff --git a/modules/audio_processing/agc2/rnn_vad/BUILD.gn b/modules/audio_processing/agc2/rnn_vad/BUILD.gn
index e05dcab..395e522 100644
--- a/modules/audio_processing/agc2/rnn_vad/BUILD.gn
+++ b/modules/audio_processing/agc2/rnn_vad/BUILD.gn
@@ -25,12 +25,15 @@
     "pitch_search_internal.cc",
     "pitch_search_internal.h",
     "ring_buffer.h",
+    "rnn.cc",
+    "rnn.h",
     "sequence_buffer.h",
     "symmetric_matrix_buffer.h",
   ]
   deps = [
     "../../../../api:array_view",
     "../../../../rtc_base:checks",
+    "//third_party/rnnoise:rnn_vad",
   ]
 }
 
@@ -53,6 +56,8 @@
   unittest_resources = [
     "../../../../resources/audio_processing/agc2/rnn_vad/pitch_buf_24k.dat",
     "../../../../resources/audio_processing/agc2/rnn_vad/pitch_lp_res.dat",
+    "../../../../resources/audio_processing/agc2/rnn_vad/sil_features.dat",
+    "../../../../resources/audio_processing/agc2/rnn_vad/vad_prob.dat",
   ]
 
   if (is_ios) {
@@ -72,6 +77,7 @@
       "pitch_search_internal_unittest.cc",
       "pitch_search_unittest.cc",
       "ring_buffer_unittest.cc",
+      "rnn_unittest.cc",
       "sequence_buffer_unittest.cc",
       "symmetric_matrix_buffer_unittest.cc",
     ]
@@ -79,7 +85,9 @@
       ":lib",
       ":lib_test",
       "../../../../api:array_view",
+      "../../../../rtc_base:checks",
       "../../../../test:test_support",
+      "//third_party/rnnoise:rnn_vad",
     ]
     data = unittest_resources
     if (is_ios) {
diff --git a/modules/audio_processing/agc2/rnn_vad/DEPS b/modules/audio_processing/agc2/rnn_vad/DEPS
new file mode 100644
index 0000000..773c2d7
--- /dev/null
+++ b/modules/audio_processing/agc2/rnn_vad/DEPS
@@ -0,0 +1,3 @@
+include_rules = [
+  "+third_party/rnnoise",
+]
diff --git a/modules/audio_processing/agc2/rnn_vad/common.h b/modules/audio_processing/agc2/rnn_vad/common.h
index 252bf84..3af0719 100644
--- a/modules/audio_processing/agc2/rnn_vad/common.h
+++ b/modules/audio_processing/agc2/rnn_vad/common.h
@@ -43,6 +43,8 @@
 constexpr size_t kMinPitch48kHz = kMinPitch24kHz * 2;
 constexpr size_t kMaxPitch48kHz = kMaxPitch24kHz * 2;
 
+constexpr size_t kFeatureVectorSize = 42;
+
 }  // namespace rnn_vad
 }  // namespace webrtc
 
diff --git a/modules/audio_processing/agc2/rnn_vad/rnn.cc b/modules/audio_processing/agc2/rnn_vad/rnn.cc
new file mode 100644
index 0000000..f88fb75
--- /dev/null
+++ b/modules/audio_processing/agc2/rnn_vad/rnn.cc
@@ -0,0 +1,227 @@
+/*
+ *  Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "modules/audio_processing/agc2/rnn_vad/rnn.h"
+
+#include <algorithm>
+#include <array>
+#include <cmath>
+
+#include "rtc_base/checks.h"
+#include "third_party/rnnoise/src/rnn_activations.h"
+#include "third_party/rnnoise/src/rnn_vad_weights.h"
+
+namespace webrtc {
+namespace rnn_vad {
+
+using rnnoise::kWeightsScale;
+
+using rnnoise::kInputLayerInputSize;
+static_assert(kFeatureVectorSize == kInputLayerInputSize, "");
+using rnnoise::kInputDenseWeights;
+using rnnoise::kInputDenseBias;
+using rnnoise::kInputLayerOutputSize;
+static_assert(kInputLayerOutputSize <= kFullyConnectedLayersMaxUnits,
+              "Increase kFullyConnectedLayersMaxUnits.");
+
+using rnnoise::kHiddenGruRecurrentWeights;
+using rnnoise::kHiddenGruWeights;
+using rnnoise::kHiddenGruBias;
+using rnnoise::kHiddenLayerOutputSize;
+static_assert(kHiddenLayerOutputSize <= kRecurrentLayersMaxUnits,
+              "Increase kRecurrentLayersMaxUnits.");
+
+using rnnoise::kOutputDenseWeights;
+using rnnoise::kOutputDenseBias;
+using rnnoise::kOutputLayerOutputSize;
+static_assert(kOutputLayerOutputSize <= kFullyConnectedLayersMaxUnits,
+              "Increase kFullyConnectedLayersMaxUnits.");
+
+using rnnoise::RectifiedLinearUnit;
+using rnnoise::SigmoidApproximated;
+using rnnoise::TansigApproximated;
+
+FullyConnectedLayer::FullyConnectedLayer(
+    const size_t input_size,
+    const size_t output_size,
+    const rtc::ArrayView<const int8_t> bias,
+    const rtc::ArrayView<const int8_t> weights,
+    float (*const activation_function)(float))
+    : input_size_(input_size),
+      output_size_(output_size),
+      bias_(bias),
+      weights_(weights),
+      activation_function_(activation_function) {
+  RTC_DCHECK_LE(output_size_, kFullyConnectedLayersMaxUnits)
+      << "Static over-allocation of fully-connected layers output vectors is "
+         "not sufficient.";
+  RTC_DCHECK_EQ(output_size_, bias_.size())
+      << "Mismatching output size and bias terms array size.";
+  RTC_DCHECK_EQ(input_size_ * output_size_, weights_.size())
+      << "Mismatching input-output size and weight coefficients array size.";
+}
+
+FullyConnectedLayer::~FullyConnectedLayer() = default;
+
+rtc::ArrayView<const float> FullyConnectedLayer::GetOutput() const {
+  return rtc::ArrayView<const float>(output_.data(), output_size_);
+}
+
+void FullyConnectedLayer::ComputeOutput(rtc::ArrayView<const float> input) {
+  // TODO(bugs.chromium.org/9076): Optimize using SSE/AVX fused multiply-add
+  // operations.
+  for (size_t o = 0; o < output_size_; ++o) {
+    output_[o] = bias_[o];
+    // TODO(bugs.chromium.org/9076): Benchmark how different layouts for
+    // |weights_| change the performance across different platforms.
+    for (size_t i = 0; i < input_size_; ++i) {
+      output_[o] += input[i] * weights_[i * output_size_ + o];
+    }
+    output_[o] = (*activation_function_)(kWeightsScale * output_[o]);
+  }
+}
+
+GatedRecurrentLayer::GatedRecurrentLayer(
+    const size_t input_size,
+    const size_t output_size,
+    const rtc::ArrayView<const int8_t> bias,
+    const rtc::ArrayView<const int8_t> weights,
+    const rtc::ArrayView<const int8_t> recurrent_weights,
+    float (*const activation_function)(float))
+    : input_size_(input_size),
+      output_size_(output_size),
+      bias_(bias),
+      weights_(weights),
+      recurrent_weights_(recurrent_weights),
+      activation_function_(activation_function) {
+  RTC_DCHECK_LE(output_size_, kRecurrentLayersMaxUnits)
+      << "Static over-allocation of recurrent layers state vectors is not "
+      << "sufficient.";
+  RTC_DCHECK_EQ(3 * output_size_, bias_.size())
+      << "Mismatching output size and bias terms array size.";
+  RTC_DCHECK_EQ(3 * input_size_ * output_size_, weights_.size())
+      << "Mismatching input-output size and weight coefficients array size.";
+  RTC_DCHECK_EQ(3 * input_size_ * output_size_, recurrent_weights_.size())
+      << "Mismatching input-output size and recurrent weight coefficients array"
+      << " size.";
+  Reset();
+}
+
+GatedRecurrentLayer::~GatedRecurrentLayer() = default;
+
+rtc::ArrayView<const float> GatedRecurrentLayer::GetOutput() const {
+  return rtc::ArrayView<const float>(state_.data(), output_size_);
+}
+
+void GatedRecurrentLayer::Reset() {
+  state_.fill(0.f);
+}
+
+void GatedRecurrentLayer::ComputeOutput(rtc::ArrayView<const float> input) {
+  // TODO(bugs.chromium.org/9076): Optimize using SSE/AVX fused multiply-add
+  // operations.
+  // Stride and offset used to read parameter arrays.
+  const size_t stride = 3 * output_size_;
+  size_t offset = 0;
+
+  // Compute update gates.
+  std::array<float, kRecurrentLayersMaxUnits> update;
+  for (size_t o = 0; o < output_size_; ++o) {
+    update[o] = bias_[o];
+    // TODO(bugs.chromium.org/9076): Benchmark how different layouts for
+    // |weights_| and |recurrent_weights_| change the performance across
+    // different platforms.
+    for (size_t i = 0; i < input_size_; ++i) {  // Add input.
+      update[o] += input[i] * weights_[i * stride + o];
+    }
+    for (size_t s = 0; s < output_size_; ++s) {
+      update[o] += state_[s] * recurrent_weights_[s * stride + o];
+    }  // Add state.
+    update[o] = SigmoidApproximated(kWeightsScale * update[o]);
+  }
+
+  // Compute reset gates.
+  offset += output_size_;
+  std::array<float, kRecurrentLayersMaxUnits> reset;
+  for (size_t o = 0; o < output_size_; ++o) {
+    reset[o] = bias_[offset + o];
+    for (size_t i = 0; i < input_size_; ++i) {  // Add input.
+      reset[o] += input[i] * weights_[offset + i * stride + o];
+    }
+    for (size_t s = 0; s < output_size_; ++s) {  // Add state.
+      reset[o] += state_[s] * recurrent_weights_[offset + s * stride + o];
+    }
+    reset[o] = SigmoidApproximated(kWeightsScale * reset[o]);
+  }
+
+  // Compute output.
+  offset += output_size_;
+  std::array<float, kRecurrentLayersMaxUnits> output;
+  for (size_t o = 0; o < output_size_; ++o) {
+    output[o] = bias_[offset + o];
+    for (size_t i = 0; i < input_size_; ++i) {  // Add input.
+      output[o] += input[i] * weights_[offset + i * stride + o];
+    }
+    for (size_t s = 0; s < output_size_;
+         ++s) {  // Add state through reset gates.
+      output[o] +=
+          state_[s] * recurrent_weights_[offset + s * stride + o] * reset[s];
+    }
+    output[o] = (*activation_function_)(kWeightsScale * output[o]);
+    // Update output through the update gates.
+    output[o] = update[o] * state_[o] + (1.f - update[o]) * output[o];
+  }
+
+  // Update the state. Not done in the previous loop since that would pollute
+  // the current state and lead to incorrect output values.
+  std::copy(output.begin(), output.end(), state_.begin());
+}
+
+RnnBasedVad::RnnBasedVad()
+    : input_layer_(kInputLayerInputSize,
+                   kInputLayerOutputSize,
+                   kInputDenseBias,
+                   kInputDenseWeights,
+                   TansigApproximated),
+      hidden_layer_(kInputLayerOutputSize,
+                    kHiddenLayerOutputSize,
+                    kHiddenGruBias,
+                    kHiddenGruWeights,
+                    kHiddenGruRecurrentWeights,
+                    RectifiedLinearUnit),
+      output_layer_(kHiddenLayerOutputSize,
+                    kOutputLayerOutputSize,
+                    kOutputDenseBias,
+                    kOutputDenseWeights,
+                    SigmoidApproximated) {
+  // Input-output chaining size checks.
+  RTC_DCHECK_EQ(input_layer_.output_size(), hidden_layer_.input_size())
+      << "The input and the hidden layers sizes do not match.";
+  RTC_DCHECK_EQ(hidden_layer_.output_size(), output_layer_.input_size())
+      << "The hidden and the output layers sizes do not match.";
+}
+
+RnnBasedVad::~RnnBasedVad() = default;
+
+void RnnBasedVad::Reset() {
+  hidden_layer_.Reset();
+}
+
+void RnnBasedVad::ComputeVadProbability(
+    rtc::ArrayView<const float, kFeatureVectorSize> feature_vector) {
+  input_layer_.ComputeOutput(feature_vector);
+  hidden_layer_.ComputeOutput(input_layer_.GetOutput());
+  output_layer_.ComputeOutput(hidden_layer_.GetOutput());
+  const auto vad_output = output_layer_.GetOutput();
+  vad_probability_ = vad_output[0];
+}
+
+}  // namespace rnn_vad
+}  // namespace webrtc
diff --git a/modules/audio_processing/agc2/rnn_vad/rnn.h b/modules/audio_processing/agc2/rnn_vad/rnn.h
new file mode 100644
index 0000000..81ab87e
--- /dev/null
+++ b/modules/audio_processing/agc2/rnn_vad/rnn.h
@@ -0,0 +1,116 @@
+/*
+ *  Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef MODULES_AUDIO_PROCESSING_AGC2_RNN_VAD_RNN_H_
+#define MODULES_AUDIO_PROCESSING_AGC2_RNN_VAD_RNN_H_
+
+#include <array>
+
+#include "api/array_view.h"
+#include "modules/audio_processing/agc2/rnn_vad/common.h"
+
+namespace webrtc {
+namespace rnn_vad {
+
+// Maximum number of units for a fully-connected layer. This value is used to
+// over-allocate space for fully-connected layers output vectors (implemented as
+// std::array). The value should equal the number of units of the largest
+// fully-connected layer.
+constexpr size_t kFullyConnectedLayersMaxUnits = 24;
+
+// Maximum number of units for a recurrent layer. This value is used to
+// over-allocate space for recurrent layers state vectors (implemented as
+// std::array). The value should equal the number of units of the largest
+// recurrent layer.
+constexpr size_t kRecurrentLayersMaxUnits = 24;
+
+// Fully-connected layer.
+class FullyConnectedLayer {
+ public:
+  FullyConnectedLayer(const size_t input_size,
+                      const size_t output_size,
+                      const rtc::ArrayView<const int8_t> bias,
+                      const rtc::ArrayView<const int8_t> weights,
+                      float (*const activation_function)(float));
+  FullyConnectedLayer(const FullyConnectedLayer&) = delete;
+  FullyConnectedLayer& operator=(const FullyConnectedLayer&) = delete;
+  ~FullyConnectedLayer();
+  size_t input_size() const { return input_size_; }
+  size_t output_size() const { return output_size_; }
+  rtc::ArrayView<const float> GetOutput() const;
+  // Computes the fully-connected layer output.
+  void ComputeOutput(rtc::ArrayView<const float> input);
+
+ private:
+  const size_t input_size_;
+  const size_t output_size_;
+  const rtc::ArrayView<const int8_t> bias_;
+  const rtc::ArrayView<const int8_t> weights_;
+  float (*const activation_function_)(float);
+  // The output vector of a recurrent layer has length equal to |output_size_|.
+  // However, for efficiency, over-allocation is used.
+  std::array<float, kFullyConnectedLayersMaxUnits> output_;
+};
+
+// Recurrent layer with gated recurrent units (GRUs).
+class GatedRecurrentLayer {
+ public:
+  GatedRecurrentLayer(const size_t input_size,
+                      const size_t output_size,
+                      const rtc::ArrayView<const int8_t> bias,
+                      const rtc::ArrayView<const int8_t> weights,
+                      const rtc::ArrayView<const int8_t> recurrent_weights,
+                      float (*const activation_function)(float));
+  GatedRecurrentLayer(const GatedRecurrentLayer&) = delete;
+  GatedRecurrentLayer& operator=(const GatedRecurrentLayer&) = delete;
+  ~GatedRecurrentLayer();
+  size_t input_size() const { return input_size_; }
+  size_t output_size() const { return output_size_; }
+  rtc::ArrayView<const float> GetOutput() const;
+  void Reset();
+  // Computes the recurrent layer output and updates the status.
+  void ComputeOutput(rtc::ArrayView<const float> input);
+
+ private:
+  const size_t input_size_;
+  const size_t output_size_;
+  const rtc::ArrayView<const int8_t> bias_;
+  const rtc::ArrayView<const int8_t> weights_;
+  const rtc::ArrayView<const int8_t> recurrent_weights_;
+  float (*const activation_function_)(float);
+  // The state vector of a recurrent layer has length equal to |output_size_|.
+  // However, to avoid dynamic allocation, over-allocation is used.
+  std::array<float, kRecurrentLayersMaxUnits> state_;
+};
+
+// Recurrent network based VAD.
+class RnnBasedVad {
+ public:
+  RnnBasedVad();
+  RnnBasedVad(const RnnBasedVad&) = delete;
+  RnnBasedVad& operator=(const RnnBasedVad&) = delete;
+  ~RnnBasedVad();
+  float vad_probability() const { return vad_probability_; }
+  void Reset();
+  // Compute and returns the probability of voice (range: [0.0, 1.0]).
+  void ComputeVadProbability(
+      rtc::ArrayView<const float, kFeatureVectorSize> feature_vector);
+
+ private:
+  FullyConnectedLayer input_layer_;
+  GatedRecurrentLayer hidden_layer_;
+  FullyConnectedLayer output_layer_;
+  float vad_probability_;
+};
+
+}  // namespace rnn_vad
+}  // namespace webrtc
+
+#endif  // MODULES_AUDIO_PROCESSING_AGC2_RNN_VAD_RNN_H_
diff --git a/modules/audio_processing/agc2/rnn_vad/rnn_unittest.cc b/modules/audio_processing/agc2/rnn_vad/rnn_unittest.cc
new file mode 100644
index 0000000..d774c6d
--- /dev/null
+++ b/modules/audio_processing/agc2/rnn_vad/rnn_unittest.cc
@@ -0,0 +1,180 @@
+/*
+ *  Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <algorithm>
+#include <array>
+#include <vector>
+
+#include "modules/audio_processing/agc2/rnn_vad/rnn.h"
+#include "modules/audio_processing/agc2/rnn_vad/test_utils.h"
+#include "rtc_base/checks.h"
+#include "test/gtest.h"
+#include "third_party/rnnoise/src/rnn_activations.h"
+#include "third_party/rnnoise/src/rnn_vad_weights.h"
+
+namespace webrtc {
+namespace rnn_vad {
+namespace test {
+
+using rnnoise::RectifiedLinearUnit;
+using rnnoise::SigmoidApproximated;
+
+namespace {
+
+void TestFullyConnectedLayer(FullyConnectedLayer* fc,
+                             rtc::ArrayView<const float> input_vector,
+                             const float expected_output) {
+  RTC_CHECK(fc);
+  fc->ComputeOutput(input_vector);
+  const auto output = fc->GetOutput();
+  EXPECT_NEAR(expected_output, output[0], 3e-6f);
+}
+
+void TestGatedRecurrentLayer(
+    GatedRecurrentLayer* gru,
+    rtc::ArrayView<const float> input_sequence,
+    rtc::ArrayView<const float> expected_output_sequence) {
+  RTC_CHECK(gru);
+  auto gru_output_view = gru->GetOutput();
+  const size_t input_sequence_length =
+      rtc::CheckedDivExact(input_sequence.size(), gru->input_size());
+  const size_t output_sequence_length =
+      rtc::CheckedDivExact(expected_output_sequence.size(), gru->output_size());
+  ASSERT_EQ(input_sequence_length, output_sequence_length)
+      << "The test data length is invalid.";
+  // Feed the GRU layer and check the output at every step.
+  gru->Reset();
+  for (size_t i = 0; i < input_sequence_length; ++i) {
+    SCOPED_TRACE(i);
+    gru->ComputeOutput(
+        input_sequence.subview(i * gru->input_size(), gru->input_size()));
+    const auto expected_output = expected_output_sequence.subview(
+        i * gru->output_size(), gru->output_size());
+    ExpectNearAbsolute(expected_output, gru_output_view, 3e-6f);
+  }
+}
+
+}  // namespace
+
+// Bit-exactness check for fully connected layers.
+TEST(RnnVadTest, CheckFullyConnectedLayerOutput) {
+  const std::array<int8_t, 1> bias = {-50};
+  const std::array<int8_t, 24> weights = {
+      127,  127,  127, 127,  127,  20,  127,  -126, -126, -54, 14,  125,
+      -126, -126, 127, -125, -126, 127, -127, -127, -57,  -30, 127, 80};
+  FullyConnectedLayer fc(24, 1, bias, weights, SigmoidApproximated);
+  // Test on different inputs.
+  {
+    const std::array<float, 24> input_vector = {
+        0.f,           0.f,           0.f,          0.f,          0.f,
+        0.f,           0.215833917f,  0.290601075f, 0.238759011f, 0.244751841f,
+        0.f,           0.0461241305f, 0.106401242f, 0.223070428f, 0.630603909f,
+        0.690453172f,  0.f,           0.387645692f, 0.166913897f, 0.f,
+        0.0327451192f, 0.f,           0.136149868f, 0.446351469f};
+    TestFullyConnectedLayer(&fc, {input_vector}, 0.436567038f);
+  }
+  {
+    const std::array<float, 24> input_vector = {
+        0.592162728f,  0.529089332f,  1.18205106f,
+        1.21736848f,   0.f,           0.470851123f,
+        0.130675942f,  0.320903003f,  0.305496395f,
+        0.0571633279f, 1.57001138f,   0.0182026215f,
+        0.0977443159f, 0.347477973f,  0.493206412f,
+        0.9688586f,    0.0320267938f, 0.244722098f,
+        0.312745273f,  0.f,           0.00650715502f,
+        0.312553257f,  1.62619662f,   0.782880902f};
+    TestFullyConnectedLayer(&fc, {input_vector}, 0.874741316f);
+  }
+  {
+    const std::array<float, 24> input_vector = {
+        0.395022154f,  0.333681047f,  0.76302278f,
+        0.965480626f,  0.f,           0.941198349f,
+        0.0892967582f, 0.745046318f,  0.635769248f,
+        0.238564298f,  0.970656633f,  0.014159563f,
+        0.094203949f,  0.446816623f,  0.640755892f,
+        1.20532358f,   0.0254284926f, 0.283327013f,
+        0.726210058f,  0.0550272502f, 0.000344108557f,
+        0.369803518f,  1.56680179f,   0.997883797f};
+    TestFullyConnectedLayer(&fc, {input_vector}, 0.672785878f);
+  }
+}
+
+TEST(RnnVadTest, CheckGatedRecurrentLayer) {
+  const std::array<int8_t, 12> bias = {96,   -99, -81, -114, 49,  119,
+                                       -118, 68,  -76, 91,   121, 125};
+  const std::array<int8_t, 60> weights = {
+      124, 9,    1,    116, -66, -21, -118, -110, 104,  75,  -23,  -51,
+      -72, -111, 47,   93,  77,  -98, 41,   -8,   40,   -23, -43,  -107,
+      9,   -73,  30,   -32, -2,  64,  -26,  91,   -48,  -24, -28,  -104,
+      74,  -46,  116,  15,  32,  52,  -126, -38,  -121, 12,  -16,  110,
+      -95, 66,   -103, -35, -38, 3,   -126, -61,  28,   98,  -117, -43};
+  const std::array<int8_t, 60> recurrent_weights = {
+      -3,  87,  50,  51,  -22,  27,  -39, 62,   31,  -83, -52,  -48,
+      -6,  83,  -19, 104, 105,  48,  23,  68,   23,  40,  7,    -120,
+      64,  -62, 117, 85,  -51,  -43, 54,  -105, 120, 56,  -128, -107,
+      39,  50,  -17, -47, -117, 14,  108, 12,   -7,  -72, 103,  -87,
+      -66, 82,  84,  100, -98,  102, -49, 44,   122, 106, -20,  -69};
+  GatedRecurrentLayer gru(5, 4, bias, weights, recurrent_weights,
+                          RectifiedLinearUnit);
+  // Test on different inputs.
+  {
+    const std::array<float, 20> input_sequence = {
+        0.89395463f, 0.93224651f, 0.55788344f, 0.32341808f, 0.93355054f,
+        0.13475326f, 0.97370994f, 0.14253306f, 0.93710381f, 0.76093364f,
+        0.65780413f, 0.41657975f, 0.49403164f, 0.46843281f, 0.75138855f,
+        0.24517593f, 0.47657707f, 0.57064998f, 0.435184f,   0.19319285f};
+    const std::array<float, 16> expected_output_sequence = {
+        0.0239123f,  0.5773077f,  0.f,         0.f,
+        0.01282811f, 0.64330572f, 0.f,         0.04863098f,
+        0.00781069f, 0.75267816f, 0.f,         0.02579715f,
+        0.00471378f, 0.59162533f, 0.11087593f, 0.01334511f};
+    TestGatedRecurrentLayer(&gru, input_sequence, expected_output_sequence);
+  }
+}
+
+// TODO(bugs.webrtc.org/9076): Remove when the issue is fixed.
+// Bit-exactness test checking that precomputed frame-wise features lead to the
+// expected VAD probabilities.
+TEST(RnnVadTest, RnnBitExactness) {
+  // Init.
+  auto features_reader = CreateSilenceFlagsFeatureMatrixReader();
+  auto vad_probs_reader = CreateVadProbsReader();
+  ASSERT_EQ(features_reader.second, vad_probs_reader.second);
+  const size_t num_frames = features_reader.second;
+  // Frame-wise buffers.
+  float expected_vad_probability;
+  float is_silence;
+  std::array<float, kFeatureVectorSize> features;
+
+  // Compute VAD probability using the precomputed features.
+  RnnBasedVad vad;
+  for (size_t i = 0; i < num_frames; ++i) {
+    SCOPED_TRACE(i);
+    // Read frame data.
+    RTC_CHECK(vad_probs_reader.first->ReadValue(&expected_vad_probability));
+    // The features file also includes a silence flag for each frame.
+    RTC_CHECK(features_reader.first->ReadValue(&is_silence));
+    RTC_CHECK(
+        features_reader.first->ReadChunk({features.data(), features.size()}));
+    // Skip silent frames.
+    ASSERT_TRUE(is_silence == 0.f || is_silence == 1.f);
+    if (is_silence == 1.f) {
+      ASSERT_EQ(expected_vad_probability, 0.f);
+      continue;
+    }
+    // Compute and check VAD probability.
+    vad.ComputeVadProbability({features.data(), features.size()});
+    EXPECT_NEAR(expected_vad_probability, vad.vad_probability(), 3e-6f);
+  }
+}
+
+}  // namespace test
+}  // namespace rnn_vad
+}  // namespace webrtc
diff --git a/modules/audio_processing/agc2/rnn_vad/test_utils.cc b/modules/audio_processing/agc2/rnn_vad/test_utils.cc
index c6cf21e..ff91ef7 100644
--- a/modules/audio_processing/agc2/rnn_vad/test_utils.cc
+++ b/modules/audio_processing/agc2/rnn_vad/test_utils.cc
@@ -53,6 +53,21 @@
           rtc::CheckedDivExact(ptr->data_length(), 2 + num_lp_residual_coeffs)};
 }
 
+ReaderPairType CreateSilenceFlagsFeatureMatrixReader() {
+  auto ptr = rtc::MakeUnique<BinaryFileReader<float>>(
+      test::ResourcePath("audio_processing/agc2/rnn_vad/sil_features", "dat"),
+      42);
+  // Features (42) and silence flag.
+  return {std::move(ptr),
+          rtc::CheckedDivExact(ptr->data_length(), static_cast<size_t>(43))};
+}
+
+ReaderPairType CreateVadProbsReader() {
+  auto ptr = rtc::MakeUnique<BinaryFileReader<float>>(
+      test::ResourcePath("audio_processing/agc2/rnn_vad/vad_prob", "dat"));
+  return {std::move(ptr), ptr->data_length()};
+}
+
 }  // namespace test
 }  // namespace rnn_vad
 }  // namespace webrtc
diff --git a/modules/audio_processing/agc2/rnn_vad/test_utils.h b/modules/audio_processing/agc2/rnn_vad/test_utils.h
index 3f580ab..92d3706 100644
--- a/modules/audio_processing/agc2/rnn_vad/test_utils.h
+++ b/modules/audio_processing/agc2/rnn_vad/test_utils.h
@@ -95,6 +95,12 @@
 // and gain values.
 std::pair<std::unique_ptr<BinaryFileReader<float>>, const size_t>
 CreateLpResidualAndPitchPeriodGainReader();
+// Instance a reader for the silence flags and the feature matrix.
+std::pair<std::unique_ptr<BinaryFileReader<float>>, const size_t>
+CreateSilenceFlagsFeatureMatrixReader();
+// Instance a reader for the VAD probabilities.
+std::pair<std::unique_ptr<BinaryFileReader<float>>, const size_t>
+CreateVadProbsReader();
 
 }  // namespace test
 }  // namespace rnn_vad
diff --git a/resources/audio_processing/agc2/rnn_vad/sil_features.dat.sha1 b/resources/audio_processing/agc2/rnn_vad/sil_features.dat.sha1
new file mode 100644
index 0000000..bc591e9
--- /dev/null
+++ b/resources/audio_processing/agc2/rnn_vad/sil_features.dat.sha1
@@ -0,0 +1 @@
+e0a92782c2903be9da10385d924d34e8bf212d5e
\ No newline at end of file
diff --git a/resources/audio_processing/agc2/rnn_vad/vad_prob.dat.sha1 b/resources/audio_processing/agc2/rnn_vad/vad_prob.dat.sha1
new file mode 100644
index 0000000..1aa3bd0
--- /dev/null
+++ b/resources/audio_processing/agc2/rnn_vad/vad_prob.dat.sha1
@@ -0,0 +1 @@
+05735ede0b457318e307d12f5acfd11bbbbd0afd
\ No newline at end of file
diff --git a/tools_webrtc/libs/generate_licenses.py b/tools_webrtc/libs/generate_licenses.py
index 9bbe752..df7ad82 100755
--- a/tools_webrtc/libs/generate_licenses.py
+++ b/tools_webrtc/libs/generate_licenses.py
@@ -44,6 +44,7 @@
     'openmax_dl': ['third_party/openmax_dl/LICENSE'],
     'opus': ['third_party/opus/src/COPYING'],
     'protobuf': ['third_party/protobuf/LICENSE'],
+    'rnnoise': ['third_party/rnnoise/COPYING'],
     'usrsctp': ['third_party/usrsctp/LICENSE'],
     'webrtc': ['LICENSE', 'LICENSE_THIRD_PARTY'],
     'zlib': ['third_party/zlib/LICENSE'],