diff --git a/CHANGELOG.md b/CHANGELOG.md
index c4f4301..75b562c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,36 @@
-# Changelog
-
+# Changelog
+
+## <small>2.1.5 (2025-09-12)</small>
+
+* fix: avoid crash on ios when multiple starts are thrown ([1aa44be](https://github.com/riderodd/react-native-vosk/commit/1aa44be))
+* fix: Fix format rate error on iOS (#104) ([d8fe8b0](https://github.com/riderodd/react-native-vosk/commit/d8fe8b0)), closes [#104](https://github.com/riderodd/react-native-vosk/issues/104)
+* fix: ios pcm bitrate is now compatible with device ([3e822fe](https://github.com/riderodd/react-native-vosk/commit/3e822fe))
+* fix: yarn lockfile (#155) ([507e67d](https://github.com/riderodd/react-native-vosk/commit/507e67d)), closes [#155](https://github.com/riderodd/react-native-vosk/issues/155)
+* #67: replace file:// in name path, have a more descriptive log (#79) ([4e756da](https://github.com/riderodd/react-native-vosk/commit/4e756da)), closes [#67](https://github.com/riderodd/react-native-vosk/issues/67) [#79](https://github.com/riderodd/react-native-vosk/issues/79)
+* Add Expo support for react-native-vosk (#151) ([fec6fb9](https://github.com/riderodd/react-native-vosk/commit/fec6fb9)), closes [#151](https://github.com/riderodd/react-native-vosk/issues/151)
+* Create dependabot.yml (#71) ([8634eba](https://github.com/riderodd/react-native-vosk/commit/8634eba)), closes [#71](https://github.com/riderodd/react-native-vosk/issues/71)
+* fix crashes (#143) ([cc9ea78](https://github.com/riderodd/react-native-vosk/commit/cc9ea78)), closes [#143](https://github.com/riderodd/react-native-vosk/issues/143)
+* Fix expo ios model path (#156) ([e2877d9](https://github.com/riderodd/react-native-vosk/commit/e2877d9)), closes [#156](https://github.com/riderodd/react-native-vosk/issues/156)
+* Fix Expo plugin configuration and dependencies (#152) ([3fa62f7](https://github.com/riderodd/react-native-vosk/commit/3fa62f7)), closes [#152](https://github.com/riderodd/react-native-vosk/issues/152)
+* Fix/expo ios (#154) ([cbdf9eb](https://github.com/riderodd/react-native-vosk/commit/cbdf9eb)), closes [#154](https://github.com/riderodd/react-native-vosk/issues/154)
+* Fix/update version (#106) ([eeebd52](https://github.com/riderodd/react-native-vosk/commit/eeebd52)), closes [#106](https://github.com/riderodd/react-native-vosk/issues/106)
+* Try to make expo plugin works (#153) ([b54f844](https://github.com/riderodd/react-native-vosk/commit/b54f844)), closes [#153](https://github.com/riderodd/react-native-vosk/issues/153)
+* Update dependabot.yml (#72) ([4013cf3](https://github.com/riderodd/react-native-vosk/commit/4013cf3)), closes [#72](https://github.com/riderodd/react-native-vosk/issues/72)
+* Update issue templates (#76) ([9e4ac88](https://github.com/riderodd/react-native-vosk/commit/9e4ac88)), closes [#76](https://github.com/riderodd/react-native-vosk/issues/76)
+* Update to TurboModules with breaking changes (#150) ([4869ff0](https://github.com/riderodd/react-native-vosk/commit/4869ff0)), closes [#150](https://github.com/riderodd/react-native-vosk/issues/150)
+* Update Vosk Android library and model loading method (#149) ([3248919](https://github.com/riderodd/react-native-vosk/commit/3248919)), closes [#149](https://github.com/riderodd/react-native-vosk/issues/149)
+* docs: add Copilot Coding Agent onboarding instructions (#148) ([1881c34](https://github.com/riderodd/react-native-vosk/commit/1881c34)), closes [#148](https://github.com/riderodd/react-native-vosk/issues/148)
+* chore: release 0.3.2 (#80) ([37ec42a](https://github.com/riderodd/react-native-vosk/commit/37ec42a)), closes [#80](https://github.com/riderodd/react-native-vosk/issues/80)
+* chore: updated release it deps (#105) ([b73429c](https://github.com/riderodd/react-native-vosk/commit/b73429c)), closes [#105](https://github.com/riderodd/react-native-vosk/issues/105)
+* chore(deps-dev): bump @react-native/metro-config from 0.76.1 to 0.79.2 (#118) ([a14e95a](https://github.com/riderodd/react-native-vosk/commit/a14e95a)), closes [#118](https://github.com/riderodd/react-native-vosk/issues/118)
+
+
+### BREAKING CHANGE
+
+* the model is not searched in the android project by default anymore
+
+* chore: release 1.0.0
+
 ## <small>2.1.4 (2025-09-12)</small>
 
 * chore: add changelog.md ([12de6fa](https://github.com/riderodd/react-native-vosk/commit/12de6fa))
@@ -28,4 +59,4 @@
 
 * the model is not searched in the android project by default anymore
 
-* chore: release 1.0.0
+* chore: release 1.0.0
diff --git a/example/ios/Podfile.lock b/example/ios/Podfile.lock
index 8371da0..a7bea9c 100644
--- a/example/ios/Podfile.lock
+++ b/example/ios/Podfile.lock
@@ -1835,7 +1835,7 @@ PODS:
     - ReactCommon/turbomodule/core
     - SocketRocket
     - Yoga
-  - react-native-vosk (1.0.0):
+  - react-native-vosk (2.1.4):
     - boost
     - DoubleConversion
     - fast_float
@@ -2643,7 +2643,7 @@ SPEC CHECKSUMS:
   React-Mapbuffer: 91e0eab42a6ae7f3e34091a126d70fc53bd3823e
   React-microtasksnativemodule: 1ead4fe154df3b1ba34b5a9e35ef3c4bdfa72ccb
   react-native-safe-area-context: c6e2edd1c1da07bdce287fa9d9e60c5f7b514616
-  react-native-vosk: d6e75b7aeab1b3ce0beeee5aec2c4adaa898ef7b
+  react-native-vosk: 25bd5e5a7b3b1197ebf56c65e783c9852cc04fa8
   React-NativeModulesApple: eff2eba56030eb0d107b1642b8f853bc36a833ac
   React-oscompat: b12c633e9c00f1f99467b1e0e0b8038895dae436
   React-perflogger: 58d12c4e5df1403030c97b9c621375c312cca454
diff --git a/ios/Vosk.mm b/ios/Vosk.mm
index dda9f9a..f80a2b8 100644
--- a/ios/Vosk.mm
+++ b/ios/Vosk.mm
@@ -1,268 +1,382 @@
-#import "Vosk.h"
-#import <Foundation/Foundation.h>
-#import <AVFoundation/AVFoundation.h>
-#import "RNVoskModel.h"
-#import "Vosk-API.h"
-
-// Specific key to detect execution on the processing queue
-static void *kVoskProcessingQueueKey = &kVoskProcessingQueueKey;
-
-@implementation Vosk {
-  // État interne migré depuis Swift
-  RNVoskModel *_Nullable _currentModel;
-  VoskRecognizer *_Nullable _recognizer;
-  AVAudioEngine *_audioEngine;
-  AVAudioInputNode *_inputNode;
-  AVAudioFormat *_formatInput;
-  dispatch_queue_t _processingQueue;
-  NSString *_Nullable _lastPartial;
-  dispatch_source_t _Nullable _timeoutSource; // high-performance GCD timer
-  // removed _hasListener management: we always emit
-  BOOL _isRunning; // protects use of recognizer after stop
-} RCT_EXPORT_MODULE()
-
-
-- (instancetype)init {
-  if ((self = [super init])) {
-  _processingQueue = dispatch_queue_create("recognizerQueue", DISPATCH_QUEUE_SERIAL);
-  dispatch_queue_set_specific(_processingQueue, kVoskProcessingQueueKey, kVoskProcessingQueueKey, NULL);
-    _audioEngine = [AVAudioEngine new];
-    _inputNode = _audioEngine.inputNode;
-    _formatInput = [_inputNode inputFormatForBus:0];
-    _recognizer = NULL;
-    _currentModel = nil;
-    _lastPartial = nil;
-  _timeoutSource = nil;
-    _isRunning = NO;
-  }
-  return self;
-}
-
-- (void)dealloc {
-  if (_recognizer) {
-    vosk_recognizer_free(_recognizer);
-    _recognizer = NULL;
-  }
-}
-
-- (NSArray<NSString *> *)supportedEvents {
-  return @[ @"onError", @"onResult", @"onFinalResult", @"onPartialResult", @"onTimeout" ];
-}
-
-- (void)loadModel:(nonnull NSString *)path resolve:(nonnull RCTPromiseResolveBlock)resolve reject:(nonnull RCTPromiseRejectBlock)reject { 
-  // Unload the current model if any
-  _currentModel = nil;
-  NSError *err = nil;
-  RNVoskModel *model = [[RNVoskModel alloc] initWithName:path error:&err];
-  if (model && !err) {
-    _currentModel = model;
-    resolve(nil);
-  } else {
-    reject(@"loadModel", err.localizedDescription ?: @"Failed to load model", err);
-  }
-}
-
-// Make the options nullable to avoid issues with codegen and optional fields
-- (void)start:(JS::NativeVosk::VoskOptions *_Nullable)options resolve:(nonnull RCTPromiseResolveBlock)resolve reject:(nonnull RCTPromiseRejectBlock)reject {
-  if (_currentModel == nil) {
-    reject(@"start", @"Model not loaded", nil);
-    return;
-  }
-
-  AVAudioSession *audioSession = [AVAudioSession sharedInstance];
-
-  // Extract options (grammar, timeout) from the codegen structure
-  NSArray<NSString *> *grammar = nil;
-  double timeoutMs = -1;
-
-  if (options && options->grammar()) {
-    auto gVec = options->grammar();
-    if (gVec) {
-      NSMutableArray<NSString *> *tmp = [NSMutableArray new];
-      size_t count = gVec->size();
-      for (size_t i = 0; i < count; ++i) {
-        NSString *s = gVec->at(static_cast<int>(i));
-        if (s) [tmp addObject:s];
-      }
-      grammar = tmp.count > 0 ? tmp : nil;
-    }
-  }
-
-  if (options && options->timeout()) {
-    timeoutMs = *(options->timeout());
-  }
-
-  CFAbsoluteTime tStart = CFAbsoluteTimeGetCurrent();
-  @try {
-    // Configure audio session
-    NSError *err = nil;
-    if (@available(iOS 10.0, *)) {
-      [audioSession setCategory:AVAudioSessionCategoryRecord mode:AVAudioSessionModeMeasurement options:0 error:&err];
-    } else {
-      [audioSession setCategory:AVAudioSessionCategoryRecord error:&err];
-    }
-    if (err) { @throw [NSException exceptionWithName:@"AudioSession" reason:err.localizedDescription userInfo:nil]; }
-    [audioSession setActive:YES error:&err];
-    if (err) { @throw [NSException exceptionWithName:@"AudioSession" reason:err.localizedDescription userInfo:nil]; }
-    _formatInput = [_inputNode inputFormatForBus:0];
-    const double sampleRate = (_formatInput.sampleRate > 0) ? _formatInput.sampleRate : 16000.0;
-    const uint32_t channelCount = 1;
-    const AVAudioCommonFormat commonFmt = AVAudioPCMFormatInt16;
-    const AVAudioFrameCount bufferSize = (AVAudioFrameCount)(sampleRate / 10.0);
-    _isRunning = YES;
-
-    __weak __typeof(self) weakSelf = self;
-
-    // Asynchronous creation of recognizer (expensive) on processing queue
-    dispatch_async(_processingQueue, ^{
-      __strong __typeof(self) self = weakSelf; if (!self || !self->_isRunning) return;
-      CFAbsoluteTime tRec0 = CFAbsoluteTimeGetCurrent();
-      if (grammar != nil && grammar.count > 0) {
-        NSError *jsonErr = nil;
-        NSData *jsonGrammar = [NSJSONSerialization dataWithJSONObject:grammar options:0 error:&jsonErr];
-        if (jsonGrammar && !jsonErr) {
-          std::string grammarStd((const char *)[jsonGrammar bytes], [jsonGrammar length]);
-          self->_recognizer = vosk_recognizer_new_grm(self->_currentModel.model, (float)sampleRate, grammarStd.c_str());
-        } else {
-          self->_recognizer = vosk_recognizer_new(self->_currentModel.model, (float)sampleRate);
-        }
-      } else {
-        self->_recognizer = vosk_recognizer_new(self->_currentModel.model, (float)sampleRate);
-      }
-      CFAbsoluteTime tRec1 = CFAbsoluteTimeGetCurrent();
-      NSLog(@"[Vosk] Recognizer init: %.2f ms", (tRec1 - tRec0)*1000.0);
-      if (!self->_recognizer || !self->_isRunning) return;
-
-      dispatch_async(dispatch_get_main_queue(), ^{
-        if (!self->_isRunning) return;
-        AVAudioFormat *formatPcm = [[AVAudioFormat alloc] initWithCommonFormat:commonFmt sampleRate:sampleRate channels:channelCount interleaved:YES];
-        if (!formatPcm) return;
-        __weak __typeof(self) weakSelfTap = self;
-        [self->_inputNode installTapOnBus:0 bufferSize:bufferSize format:formatPcm block:^(AVAudioPCMBuffer *buffer, AVAudioTime *when) {
-          __strong __typeof(self) self = weakSelfTap; if (!self) return; if (!self->_isRunning) return;
-          dispatch_async(self->_processingQueue, ^{
-            if (!self->_isRunning) return; VoskRecognizer *recognizer = self->_recognizer; if (!recognizer) return;
-            if (!buffer.int16ChannelData || !buffer.int16ChannelData[0]) return;
-            int dataLen = (int)(buffer.frameLength * 2);
-            int endOfSpeech = vosk_recognizer_accept_waveform(recognizer, (const char *)buffer.int16ChannelData[0], (int32_t)dataLen);
-            const char *cstr = endOfSpeech == 1 ? vosk_recognizer_result(recognizer) : vosk_recognizer_partial_result(recognizer);
-            NSString *json = cstr ? [NSString stringWithUTF8String:cstr] : nil;
-            dispatch_async(dispatch_get_main_queue(), ^{
-              if (!json) return; NSData *data = [json dataUsingEncoding:NSUTF8StringEncoding]; NSDictionary *parsed = data ? [NSJSONSerialization JSONObjectWithData:data options:0 error:nil] : nil;
-              if (![parsed isKindOfClass:[NSDictionary class]]) { if (endOfSpeech == 1) { [self emitOnResult:json]; } else { [self emitOnPartialResult:json]; } return; }
-              NSString *text = parsed[@"text"]; NSString *partial = parsed[@"partial"]; if (endOfSpeech == 1) { if (text.length > 0) [self emitOnResult:text]; } else { if (partial.length > 0 && (!self->_lastPartial || ![self->_lastPartial isEqualToString:partial])) { [self emitOnPartialResult:partial]; } self->_lastPartial = partial ?: self->_lastPartial; }
-            });
-          });
-        }];
-        [self->_audioEngine prepare];
-        [audioSession requestRecordPermission:^(BOOL granted) {
-          if (!granted) return; NSError *startErr = nil; [self->_audioEngine startAndReturnError:&startErr];
-        }];
-      });
-    });
-
-    // GCD timer for timeout
-    if (timeoutMs >= 0) {
-      _timeoutSource = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, _processingQueue);
-      if (_timeoutSource) {
-        uint64_t delayNs = (uint64_t)(timeoutMs * 1000000.0); // ms -> ns
-        dispatch_source_set_timer(_timeoutSource, dispatch_time(DISPATCH_TIME_NOW, delayNs), DISPATCH_TIME_FOREVER, 5 * NSEC_PER_MSEC);
-        __weak __typeof(self) weakSelfTimeout = self;
-        dispatch_source_set_event_handler(_timeoutSource, ^{
-          __strong __typeof(self) selfT = weakSelfTimeout; if (!selfT) return; if (!selfT->_isRunning) return;
-          [selfT stopInternalWithoutEvents:YES];
-          dispatch_async(dispatch_get_main_queue(), ^{ [selfT emitOnTimeout]; });
-        });
-        dispatch_resume(_timeoutSource);
-      }
-    }
-
-    CFAbsoluteTime tEnd = CFAbsoluteTimeGetCurrent();
-    NSLog(@"[Vosk] start() sync phase: %.2f ms", (tEnd - tStart)*1000.0);
-    // early resolve to release JS quickly
-    resolve(nil);
-    return;
-  } @catch (NSException *ex) {
-    // Log and emit error
-    NSLog(@"Error starting audio engine: %@", ex.reason);
-  [self emitOnError:[NSString stringWithFormat:@"Unable to start AVAudioEngine %@", ex.reason ?: @""]];
-    if (_recognizer) {
-      vosk_recognizer_free(_recognizer);
-      _recognizer = NULL;
-    }
-    reject(@"start", ex.reason ?: @"Unknown error", nil);
-  }
-}
-
-- (void)stop { 
-  if (!_isRunning) return; // idempotent
-  [self stopInternalWithoutEvents:NO];
-}
-
-- (void)unload { 
-  if (_isRunning) {
-    [self stopInternalWithoutEvents:NO];
-  }
-  _currentModel = nil;
-}
-
-- (void)addListener:(nonnull NSString *)eventType { 
-  
-}
-
-
-- (void)removeListeners:(double)count { 
-  
-}
-
-
-- (std::shared_ptr<facebook::react::TurboModule>)getTurboModule:(const facebook::react::ObjCTurboModule::InitParams &)params { 
-  return std::make_shared<facebook::react::NativeVoskSpecJSI>(params);
-}
-
-// Internal cleanup
-- (void)stopInternalWithoutEvents:(BOOL)withoutEvents {
-  @try {
-    [_inputNode removeTapOnBus:0];
-  } @catch (...) {}
-
-  if (!_isRunning) {
-    // already stopped
-  }
-  _isRunning = NO; // prevents new buffers from being processed
-
-  if (_audioEngine.isRunning) {
-    [_audioEngine stop];
-    if (!withoutEvents) {
-      [self emitOnFinalResult:_lastPartial];
-    }
-    _lastPartial = nil;
-  }
-  // Recognizer cleanup after draining processing queue, without deadlock if already on the queue
-  if (dispatch_get_specific(kVoskProcessingQueueKey)) {
-    if (_recognizer) {
-      vosk_recognizer_free(_recognizer);
-      _recognizer = NULL;
-    }
-  } else {
-    dispatch_sync(_processingQueue, ^{
-      if (self->_recognizer) {
-        vosk_recognizer_free(self->_recognizer);
-        self->_recognizer = NULL;
-      }
-    });
-  }
-  if (_timeoutSource) { dispatch_source_cancel(_timeoutSource); _timeoutSource = nil; }
-
-  AVAudioSession *audioSession = [AVAudioSession sharedInstance];
-  NSError *err = nil;
-  if (@available(iOS 10.0, *)) {
-    [audioSession setCategory:AVAudioSessionCategoryPlayback mode:AVAudioSessionModeDefault options:0 error:&err];
-  } else {
-    [audioSession setCategory:AVAudioSessionCategoryPlayback error:&err];
-  }
-  [audioSession setActive:NO withOptions:AVAudioSessionSetActiveOptionNotifyOthersOnDeactivation error:&err];
-}
-
-@end
-
+#import "Vosk.h"
+#import "RNVoskModel.h"
+#import "Vosk-API.h"
+#import <AVFoundation/AVFoundation.h>
+#import <Foundation/Foundation.h>
+#include <vector> // added for float32 -> int16 conversion
+
+// Specific key to detect execution on the processing queue
+static void *kVoskProcessingQueueKey = &kVoskProcessingQueueKey;
+
+@implementation Vosk {
+  // État interne migré depuis Swift
+  RNVoskModel *_Nullable _currentModel;
+  VoskRecognizer *_Nullable _recognizer;
+  AVAudioEngine *_audioEngine;
+  AVAudioInputNode *_inputNode;
+  AVAudioFormat *_formatInput;
+  dispatch_queue_t _processingQueue;
+  NSString *_Nullable _lastPartial;
+  dispatch_source_t _Nullable _timeoutSource; // high-performance GCD timer
+  // removed _hasListener management: we always emit
+  BOOL _isRunning; // protects use of recognizer after stop
+  BOOL _isStarting; // prevents concurrent start
+  BOOL _tapInstalled; // track tap installation
+}
+RCT_EXPORT_MODULE()
+
+- (instancetype)init {
+  if ((self = [super init])) {
+    _processingQueue =
+        dispatch_queue_create("recognizerQueue", DISPATCH_QUEUE_SERIAL);
+    dispatch_queue_set_specific(_processingQueue, kVoskProcessingQueueKey,
+                                kVoskProcessingQueueKey, NULL);
+    _audioEngine = [AVAudioEngine new];
+    _inputNode = _audioEngine.inputNode;
+    _formatInput = [_inputNode inputFormatForBus:0];
+    _recognizer = NULL;
+    _currentModel = nil;
+    _lastPartial = nil;
+    _timeoutSource = nil;
+    _isRunning = NO;
+    _isStarting = NO;
+    _tapInstalled = NO;
+  }
+  return self;
+}
+
+- (void)dealloc {
+  if (_recognizer) {
+    vosk_recognizer_free(_recognizer);
+    _recognizer = NULL;
+  }
+}
+
+- (NSArray<NSString *> *)supportedEvents {
+  return @[
+    @"onError", @"onResult", @"onFinalResult", @"onPartialResult", @"onTimeout"
+  ];
+}
+
+- (void)loadModel:(nonnull NSString *)path
+          resolve:(nonnull RCTPromiseResolveBlock)resolve
+           reject:(nonnull RCTPromiseRejectBlock)reject {
+  // Unload the current model if any
+  _currentModel = nil;
+  NSError *err = nil;
+  RNVoskModel *model = [[RNVoskModel alloc] initWithName:path error:&err];
+  if (model && !err) {
+    _currentModel = model;
+    resolve(nil);
+  } else {
+    reject(@"loadModel", err.localizedDescription ?: @"Failed to load model",
+           err);
+  }
+}
+
+// Make the options nullable to avoid issues with codegen and optional fields
+- (void)start:(JS::NativeVosk::VoskOptions *_Nullable)options
+      resolve:(nonnull RCTPromiseResolveBlock)resolve
+       reject:(nonnull RCTPromiseRejectBlock)reject {
+  if (_currentModel == nil) {
+    reject(@"start", @"Model not loaded", nil);
+    return;
+  }
+  if (_isStarting || _isRunning) {
+    reject(@"start", @"Recognizer already starting or running", nil);
+    return;
+  }
+  _isStarting = YES;
+  AVAudioSession *audioSession = [AVAudioSession sharedInstance];
+
+  // Extract options (grammar, timeout) from the codegen structure
+  NSArray<NSString *> *grammar = nil;
+  double timeoutMs = -1;
+  if (options && options->grammar()) {
+    auto gVec = options->grammar();
+    if (gVec) {
+      NSMutableArray<NSString *> *tmp = [NSMutableArray new];
+      size_t count = gVec->size();
+      for (size_t i = 0; i < count; ++i) {
+        NSString *s = gVec->at(static_cast<int>(i));
+        if (s) [tmp addObject:s];
+      }
+      grammar = tmp.count > 0 ? tmp : nil;
+    }
+  }
+  if (options && options->timeout()) {
+    timeoutMs = *(options->timeout());
+  }
+
+  CFAbsoluteTime tStart = CFAbsoluteTimeGetCurrent();
+  // Configure session category first (allowed even before permission)
+  NSError *catErr = nil;
+  @try {
+    if (@available(iOS 10.0, *)) {
+      [audioSession setCategory:AVAudioSessionCategoryRecord
+                           mode:AVAudioSessionModeMeasurement
+                        options:0
+                          error:&catErr];
+    } else {
+      [audioSession setCategory:AVAudioSessionCategoryRecord error:&catErr];
+    }
+    if (catErr) {
+      NSString *msg = [NSString stringWithFormat:@"Audio session category error: %@", catErr.localizedDescription];
+      [self emitOnError:msg];
+      reject(@"start", msg, catErr);
+      return;
+    }
+  } @catch (NSException *ex) {
+    NSString *msg = [NSString stringWithFormat:@"Exception setting category: %@", ex.reason ?: @"unknown"]; 
+    [self emitOnError:msg];
+    reject(@"start", msg, nil);
+    return;
+  }
+
+  // Request permission BEFORE doing heavy work
+  [audioSession requestRecordPermission:^(BOOL granted) {
+    if (!granted) {
+      dispatch_async(dispatch_get_main_queue(), ^{
+        NSString *msg = @"Microphone permission denied";
+        [self emitOnError:msg];
+        reject(@"start", msg, nil);
+      });
+      return;
+    }
+    dispatch_async(dispatch_get_main_queue(), ^{ // proceed on main
+      NSError *actErr = nil;
+      if (![audioSession setActive:YES error:&actErr]) {
+        NSString *msg = [NSString stringWithFormat:@"Failed to activate audio session: %@", actErr.localizedDescription];
+        [self emitOnError:msg];
+        reject(@"start", msg, actErr);
+        return;
+      }
+
+  self->_formatInput = [self->_inputNode inputFormatForBus:0];
+      const double sampleRate = (self->_formatInput.sampleRate > 0) ? self->_formatInput.sampleRate : 16000.0;
+      const AVAudioFrameCount bufferSize = (AVAudioFrameCount)(sampleRate / 10.0);
+      self->_isRunning = YES;
+
+      __weak __typeof(self) weakSelf = self;
+      dispatch_async(self->_processingQueue, ^{ // recognizer init off main
+        __strong __typeof(self) self = weakSelf;
+        if (!self || !self->_isRunning) return;
+        CFAbsoluteTime tRec0 = CFAbsoluteTimeGetCurrent();
+        if (grammar != nil && grammar.count > 0) {
+          NSError *jsonErr = nil;
+            NSData *jsonGrammar = [NSJSONSerialization dataWithJSONObject:grammar options:0 error:&jsonErr];
+          if (jsonGrammar && !jsonErr) {
+            std::string grammarStd((const char *)[jsonGrammar bytes], [jsonGrammar length]);
+            self->_recognizer = vosk_recognizer_new_grm(self->_currentModel.model, (float)sampleRate, grammarStd.c_str());
+          } else {
+            self->_recognizer = vosk_recognizer_new(self->_currentModel.model, (float)sampleRate);
+          }
+        } else {
+          self->_recognizer = vosk_recognizer_new(self->_currentModel.model, (float)sampleRate);
+        }
+        CFAbsoluteTime tRec1 = CFAbsoluteTimeGetCurrent();
+        NSLog(@"[Vosk] Recognizer init: %.2f ms", (tRec1 - tRec0) * 1000.0);
+        if (!self->_recognizer) {
+          dispatch_async(dispatch_get_main_queue(), ^{
+            [self emitOnError:@"Recognizer initialization failed (null)"];
+            reject(@"start", @"Recognizer initialization failed", nil);
+            self->_isStarting = NO;
+          });
+          return; // promise already resolved later below or not? -> we won't resolve now
+        }
+
+        dispatch_async(dispatch_get_main_queue(), ^{ // install tap + start engine
+          if (!self->_isRunning) { self->_isStarting = NO; return; }
+          if (self->_tapInstalled) {
+            NSLog(@"[Vosk] Tap already installed – skipping reinstall");
+            self->_isStarting = NO;
+            resolve(nil);
+            return;
+          }
+          AVAudioFormat *tapFormat = [self->_inputNode inputFormatForBus:0];
+          if (!tapFormat) {
+            [self emitOnError:@"tapFormat is nil; cannot install tap"]; 
+            [self stopInternalWithoutEvents:YES];
+            self->_isStarting = NO;
+            return;
+          }
+          BOOL invalidSR = tapFormat.sampleRate <= 0.0;
+          BOOL invalidCh = tapFormat.channelCount == 0;
+          if (invalidSR || invalidCh) {
+            NSLog(@"[Vosk] Invalid input format (sr=%.1f ch=%u) -> using nil format", tapFormat.sampleRate, (unsigned int)tapFormat.channelCount);
+            tapFormat = nil;
+          }
+          __weak __typeof(self) weakSelfTap = self;
+          @try {
+            [self->_inputNode installTapOnBus:0 bufferSize:bufferSize format:tapFormat block:^(AVAudioPCMBuffer *buffer, AVAudioTime *when) {
+            __strong __typeof(self) self = weakSelfTap; if (!self) return; if (!self->_isRunning) return;
+            dispatch_async(self->_processingQueue, ^{
+              if (!self->_isRunning) return; VoskRecognizer *recognizer = self->_recognizer; if (!recognizer) return;
+              AVAudioFrameCount frames = buffer.frameLength; if (frames == 0) return;
+              int accepted = 0;
+              if (buffer.int16ChannelData && buffer.int16ChannelData[0]) {
+                int dataLen = (int)(frames * sizeof(int16_t));
+                accepted = vosk_recognizer_accept_waveform(recognizer, (const char *)buffer.int16ChannelData[0], (int32_t)dataLen);
+              } else if (buffer.floatChannelData && buffer.floatChannelData[0]) {
+                float *const *floatChannels = buffer.floatChannelData; std::vector<int16_t> pcm16; pcm16.resize(frames); float *ch0 = floatChannels[0];
+                for (AVAudioFrameCount i = 0; i < frames; ++i) { float s = ch0[i]; if (s > 1.f) s = 1.f; else if (s < -1.f) s = -1.f; pcm16[i] = (int16_t)lrintf(s * 32767.f); }
+                int dataLen = (int)(frames * sizeof(int16_t));
+                accepted = vosk_recognizer_accept_waveform(recognizer, (const char *)pcm16.data(), (int32_t)dataLen);
+              } else { return; }
+
+              const char *cstr = NULL;
+              BOOL isFinal = NO;
+              if (accepted) { // end of utterance recognized per Vosk API
+                cstr = vosk_recognizer_result(recognizer);
+                isFinal = YES;
+              } else {
+                cstr = vosk_recognizer_partial_result(recognizer);
+                isFinal = NO;
+              }
+              NSString *json = cstr ? [NSString stringWithUTF8String:cstr] : nil;
+              dispatch_async(dispatch_get_main_queue(), ^{
+                if (!json) return;
+                NSData *data = [json dataUsingEncoding:NSUTF8StringEncoding];
+                NSDictionary *parsed = data ? [NSJSONSerialization JSONObjectWithData:data options:0 error:nil] : nil;
+                if (![parsed isKindOfClass:[NSDictionary class]]) {
+                  if (isFinal) { [self emitOnResult:json]; } else { [self emitOnPartialResult:json]; }
+                  return;
+                }
+                NSString *text = parsed[@"text"]; NSString *partial = parsed[@"partial"];
+                if (isFinal) {
+                  if (text.length > 0) { [self emitOnResult:text]; }
+                  self->_lastPartial = nil; // reset after final
+                } else {
+                  if (partial.length > 0 && (!self->_lastPartial || ![self->_lastPartial isEqualToString:partial])) {
+                    [self emitOnPartialResult:partial];
+                  }
+                  self->_lastPartial = partial ?: self->_lastPartial;
+                }
+              });
+            });
+            }];
+            self->_tapInstalled = YES;
+          } @catch (NSException *ex) {
+            [self emitOnError:[NSString stringWithFormat:@"Failed to install tap: %@", ex.reason ?: @"unknown"]];
+            [self stopInternalWithoutEvents:YES];
+            self->_isStarting = NO;
+            reject(@"start", @"Failed to install tap", nil);
+            return;
+          }
+          [self->_audioEngine prepare];
+          NSError *startErr = nil;
+          if (![self->_audioEngine startAndReturnError:&startErr]) {
+            NSString *msg = [NSString stringWithFormat:@"Failed to start audio engine: %@", startErr.localizedDescription];
+            [self emitOnError:msg];
+            // clean
+            [self stopInternalWithoutEvents:YES];
+            reject(@"start", msg, startErr); // still reject if not yet resolved
+            self->_isStarting = NO;
+            return;
+          }
+
+          // Timeout timer after engine started
+          if (timeoutMs >= 0) {
+            self->_timeoutSource = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, self->_processingQueue);
+            if (self->_timeoutSource) {
+              uint64_t delayNs = (uint64_t)(timeoutMs * 1000000.0);
+              __weak __typeof(self) weakSelfTimeout = self;
+              dispatch_source_set_timer(self->_timeoutSource, dispatch_time(DISPATCH_TIME_NOW, delayNs), DISPATCH_TIME_FOREVER, 5 * NSEC_PER_MSEC);
+              dispatch_source_set_event_handler(self->_timeoutSource, ^{ __strong __typeof(self) selfT = weakSelfTimeout; if (!selfT || !selfT->_isRunning) return; [selfT stopInternalWithoutEvents:YES]; dispatch_async(dispatch_get_main_queue(), ^{ [selfT emitOnTimeout]; }); });
+              dispatch_resume(self->_timeoutSource);
+            }
+          }
+
+          CFAbsoluteTime tEnd = CFAbsoluteTimeGetCurrent();
+          NSLog(@"[Vosk] start() total after permission: %.2f ms", (tEnd - tStart) * 1000.0);
+          // Resolve promise (success path)
+          self->_isStarting = NO;
+          resolve(nil);
+        });
+      });
+    });
+  }];
+}
+
+- (void)stop {
+  if (!_isRunning)
+    return; // idempotent
+  [self stopInternalWithoutEvents:NO];
+}
+
+- (void)unload {
+  if (_isRunning) {
+    [self stopInternalWithoutEvents:NO];
+  }
+  _currentModel = nil;
+}
+
+- (void)addListener:(nonnull NSString *)eventType {
+}
+
+- (void)removeListeners:(double)count {
+}
+
+- (std::shared_ptr<facebook::react::TurboModule>)getTurboModule:
+    (const facebook::react::ObjCTurboModule::InitParams &)params {
+  return std::make_shared<facebook::react::NativeVoskSpecJSI>(params);
+}
+
+// Internal cleanup
+- (void)stopInternalWithoutEvents:(BOOL)withoutEvents {
+  @try {
+    [_inputNode removeTapOnBus:0];
+  } @catch (...) {
+  }
+
+  if (!_isRunning) {
+    // already stopped
+  }
+  _isRunning = NO; // prevents new buffers from being processed
+  _isStarting = NO;
+  _tapInstalled = NO;
+
+  if (_audioEngine.isRunning) {
+    [_audioEngine stop];
+    if (!withoutEvents) {
+      [self emitOnFinalResult:_lastPartial];
+    }
+    _lastPartial = nil;
+  }
+  // Recognizer cleanup after draining processing queue, without deadlock if
+  // already on the queue
+  if (dispatch_get_specific(kVoskProcessingQueueKey)) {
+    if (_recognizer) {
+      vosk_recognizer_free(_recognizer);
+      _recognizer = NULL;
+    }
+  } else {
+    dispatch_sync(_processingQueue, ^{
+      if (self->_recognizer) {
+        vosk_recognizer_free(self->_recognizer);
+        self->_recognizer = NULL;
+      }
+    });
+  }
+  if (_timeoutSource) {
+    dispatch_source_cancel(_timeoutSource);
+    _timeoutSource = nil;
+  }
+
+  AVAudioSession *audioSession = [AVAudioSession sharedInstance];
+  NSError *err = nil;
+  if (@available(iOS 10.0, *)) {
+    [audioSession setCategory:AVAudioSessionCategoryPlayback
+                         mode:AVAudioSessionModeDefault
+                      options:0
+                        error:&err];
+  } else {
+    [audioSession setCategory:AVAudioSessionCategoryPlayback error:&err];
+  }
+  [audioSession
+        setActive:NO
+      withOptions:AVAudioSessionSetActiveOptionNotifyOthersOnDeactivation
+            error:&err];
+}
+
+@end
diff --git a/package.json b/package.json
index 15a9941..cc31733 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "react-native-vosk",
-  "version": "2.1.4",
+  "version": "2.1.5",
   "description": "Speech recognition module for react native using Vosk library",
   "types": "lib/typescript/src/index.d.ts",
   "main": "lib/commonjs/index.js",