diff --git a/src/Rtp/Decoder.cpp b/src/Rtp/Decoder.cpp
index 61cbdd7c..79ead826 100644
--- a/src/Rtp/Decoder.cpp
+++ b/src/Rtp/Decoder.cpp
@@ -125,8 +125,14 @@ void DecoderImp::onDecode(int stream, int codecid, int flags, int64_t pts, int64
     auto frame = Factory::getFrameFromPtr(codec, (char *)data, bytes, dts, pts);
     if (getTrackType(codec) != TrackVideo) {
         onFrame(stream, frame);
+        if (_last_is_keyframe && _video_merge) {
+            // 上次是关键帧，收到音频后，说明帧收齐了
+            _video_merge->flush();
+        }
         return;
     }
+    _last_is_keyframe = frame->keyFrame();
+    _video_merge = &ref.second;
     ref.second.inputFrame(frame, [this, stream, codec](uint64_t dts, uint64_t pts, const Buffer::Ptr &buffer, bool) {
         onFrame(stream, Factory::getFrameFromBuffer(codec, buffer, dts, pts));
     });
diff --git a/src/Rtp/Decoder.h b/src/Rtp/Decoder.h
index 4902157c..8a8b90a1 100644
--- a/src/Rtp/Decoder.h
+++ b/src/Rtp/Decoder.h
@@ -59,13 +59,14 @@ private:
 private:
     bool _finished = false;
     bool _have_video = false;
+    bool _last_is_keyframe = false;
     Decoder::Ptr _decoder;
     MediaSinkInterface *_sink;
-
     class FrameMergerImp : public FrameMerger {
     public:
         FrameMergerImp() : FrameMerger(FrameMerger::none) {}
     };
+    FrameMergerImp *_video_merge = nullptr;
     std::unordered_map<int, std::pair<Track::Ptr, FrameMergerImp> > _tracks;
 };