From fac0a6d7fca5e20e9d880cc39e6cae9c96cea73b Mon Sep 17 00:00:00 2001
From: Radek Czemerys <radko93@gmail.com>
Date: Mon, 6 Apr 2026 20:31:18 +0800
Subject: [PATCH 1/2] fix: add mutex to VoiceActivityDetection to prevent race
 between generate() and unload()

VoiceActivityDetection inherits from BaseModel but lacks the thread-safety
that VisionModel already provides. When generate() runs on a worker thread
(via ModelHostObject::promiseHostFunction) and unload() is called from the
JS thread, BaseModel::unload() destroys module_ mid-inference, causing
SIGILL/SIGSEGV crashes.

This applies the same pattern used by VisionModel:
- Add inference_mutex_ member
- Lock in generate() to protect forward() calls
- Override unload() to acquire the lock before BaseModel::unload()

Fixes https://github.com/software-mansion/react-native-executorch/issues/1055
---
 .../VoiceActivityDetection.cpp                      |  6 ++++++
 .../VoiceActivityDetection.h                        | 13 +++++++++++++
 2 files changed, 19 insertions(+)

diff --git a/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.cpp b/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.cpp
index a1252edfee..49971cba8c 100644
--- a/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.cpp
+++ b/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.cpp
@@ -54,8 +54,14 @@ VoiceActivityDetection::preprocess(std::span<float> waveform) const {
   return frameBuffer;
 }
 
+void VoiceActivityDetection::unload() noexcept {
+  std::scoped_lock lock(inference_mutex_);
+  BaseModel::unload();
+}
+
 std::vector<types::Segment>
 VoiceActivityDetection::generate(std::span<float> waveform) const {
+  std::scoped_lock lock(inference_mutex_);
 
   auto windowedInput = preprocess(waveform);
   auto [chunksNumber, remainder] = std::div(
diff --git a/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h b/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h
index e692889305..d2f2b1f9e6 100644
--- a/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h
+++ b/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h
@@ -5,6 +5,7 @@
 #include <executorch/extension/tensor/tensor.h>
 #include <executorch/extension/tensor/tensor_ptr.h>
 #include <executorch/runtime/core/evalue.h>
+#include <mutex>
 #include <span>
 
 #include "rnexecutorch/metaprogramming/ConstructorHelpers.h"
@@ -23,7 +24,19 @@ class VoiceActivityDetection : public BaseModel {
   [[nodiscard("Registered non-void function")]] std::vector<types::Segment>
   generate(std::span<float> waveform) const;
 
+  /**
+   * @brief Thread-safe unload that waits for any in-flight inference to
+   * complete.
+   *
+   * Mirrors VisionModel::unload(). Without this, BaseModel::unload() can
+   * destroy module_ while generate() is still calling forward() on a worker
+   * thread, causing SIGILL / SIGSEGV crashes.
+   */
+  void unload() noexcept;
+
 private:
+  mutable std::mutex inference_mutex_;
+
   std::vector<std::array<float, constants::kPaddedWindowSize>>
   preprocess(std::span<float> waveform) const;
   std::vector<types::Segment> postprocess(const std::vector<float> &scores,

From 91581fba871ae3ba3faa33cb64ac5a30f9baf699 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 7 Apr 2026 19:19:19 +0200
Subject: [PATCH 2/2] Update
 packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h

---
 .../voice_activity_detection/VoiceActivityDetection.h     | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h b/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h
index d2f2b1f9e6..c756bb6d3c 100644
--- a/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h
+++ b/packages/react-native-executorch/common/rnexecutorch/models/voice_activity_detection/VoiceActivityDetection.h
@@ -24,14 +24,6 @@ class VoiceActivityDetection : public BaseModel {
   [[nodiscard("Registered non-void function")]] std::vector<types::Segment>
   generate(std::span<float> waveform) const;
 
-  /**
-   * @brief Thread-safe unload that waits for any in-flight inference to
-   * complete.
-   *
-   * Mirrors VisionModel::unload(). Without this, BaseModel::unload() can
-   * destroy module_ while generate() is still calling forward() on a worker
-   * thread, causing SIGILL / SIGSEGV crashes.
-   */
   void unload() noexcept;
 
 private: