diff --git a/main/application.cc b/main/application.cc
index dc212fd..e78370f 100644
--- a/main/application.cc
+++ b/main/application.cc
@@ -531,10 +531,25 @@ void Application::Start() {
             opus_encoder_->Encode(std::move(data), [this](std::vector<uint8_t>&& opus) {
                 AudioStreamPacket packet;
                 packet.payload = std::move(opus);
-                packet.timestamp = last_output_timestamp_;
-                last_output_timestamp_ = 0;
-                Schedule([this, packet = std::move(packet)]() {
+                uint32_t last_output_timestamp_value = last_output_timestamp_.load();
+                {
+                    std::lock_guard<std::mutex> lock(timestamp_mutex_);
+                    if (!timestamp_queue_.empty()) {
+                        packet.timestamp = timestamp_queue_.front();
+                        timestamp_queue_.pop_front();
+                    } else {
+                        packet.timestamp = 0;
+                    }
+
+                    if (timestamp_queue_.size() > 3) { // 限制队列长度3
+                        timestamp_queue_.pop_front(); // 该包发送前先出队保持队列长度
+                        return;
+                    }
+                }
+                Schedule([this, last_output_timestamp_value, packet = std::move(packet)]() {
                     protocol_->SendAudio(packet);
+                    // ESP_LOGI(TAG, "Send %zu bytes, timestamp %lu, last_ts %lu, qsize %zu",
+                    //     packet.payload.size(), packet.timestamp, last_output_timestamp_value, timestamp_queue_.size());
                 });
             });
         });
@@ -717,7 +732,11 @@ void Application::OnAudioOutput() {
             pcm = std::move(resampled);
         }
         codec->OutputData(pcm);
-        last_output_timestamp_ = packet.timestamp;
+        {
+            std::lock_guard<std::mutex> lock(timestamp_mutex_);
+            timestamp_queue_.push_back(packet.timestamp);
+            last_output_timestamp_ = packet.timestamp;
+        }
         last_output_time_ = std::chrono::steady_clock::now();
     });
 }
@@ -816,6 +835,7 @@ void Application::SetDeviceState(DeviceState state) {
             display->SetStatus(Lang::Strings::STANDBY);
             display->SetEmotion("neutral");
             audio_processor_->Stop();
+            
 #if CONFIG_USE_WAKE_WORD_DETECT
             wake_word_detect_.StartDetection();
 #endif
@@ -824,11 +844,12 @@ void Application::SetDeviceState(DeviceState state) {
             display->SetStatus(Lang::Strings::CONNECTING);
             display->SetEmotion("neutral");
             display->SetChatMessage("system", "");
+            timestamp_queue_.clear();
+            last_output_timestamp_ = 0;
             break;
         case kDeviceStateListening:
             display->SetStatus(Lang::Strings::LISTENING);
             display->SetEmotion("neutral");
-
             // Update the IoT states before sending the start listening command
             UpdateIotStates();
 
@@ -870,7 +891,6 @@ void Application::ResetDecoder() {
     audio_decode_queue_.clear();
     audio_decode_cv_.notify_all();
     last_output_time_ = std::chrono::steady_clock::now();
-    
     auto codec = Board::GetInstance().GetAudioCodec();
     codec->EnableOutput(true);
 }
diff --git a/main/application.h b/main/application.h
index 12cd7f7..7dd3240 100644
--- a/main/application.h
+++ b/main/application.h
@@ -104,10 +104,14 @@ private:
     TaskHandle_t audio_loop_task_handle_ = nullptr;
     BackgroundTask* background_task_ = nullptr;
     std::chrono::steady_clock::time_point last_output_time_;
-    std::atomic<uint32_t> last_output_timestamp_ = 0;
     std::list<AudioStreamPacket> audio_decode_queue_;
     std::condition_variable audio_decode_cv_;
 
+    // 新增：用于维护音频包的timestamp队列
+    std::list<uint32_t> timestamp_queue_;
+    std::mutex timestamp_mutex_;
+    std::atomic<uint32_t> last_output_timestamp_ = 0;
+
     std::unique_ptr<OpusEncoderWrapper> opus_encoder_;
     std::unique_ptr<OpusDecoderWrapper> opus_decoder_;