diff --git a/README.md b/README.md
index 3b27b2f101..57a7535395 100755
--- a/README.md
+++ b/README.md
@@ -765,7 +765,9 @@ About the HLS overhead of SRS, we compare the overhead to FLV by remux the HLS t
 | 5147kbps  |   370s        |   195040      |   200280      |   2.68%       |
 | 5158kbps  |   1327s       |   835664      |   858092      |   2.68%       |
 
-The HLS overhead is calc by: (HLS - FLV) / FLV * 100%
+The HLS overhead is calc by: (HLS - FLV) / FLV * 100%.
+
+The overhead is larger than this benchmark(48kbps audio is best overhead), for we fix the [#512][bug#512].
 
 ## Architecture
 
@@ -1193,6 +1195,8 @@ Winlin
 [bug #59]: https://github.com/simple-rtmp-server/srs/issues/59
 [bug #50]: https://github.com/simple-rtmp-server/srs/issues/50
 [bug #34]: https://github.com/simple-rtmp-server/srs/issues/34
+[bug #512]: https://github.com/simple-rtmp-server/srs/issues/512
+[bug #xxxxxxxxxx]: https://github.com/simple-rtmp-server/srs/issues/xxxxxxxxxx
 
 [r2.0a2]: https://github.com/simple-rtmp-server/srs/releases/tag/v2.0-a2
 [r2.0a1]: https://github.com/simple-rtmp-server/srs/releases/tag/2.0a1
diff --git a/trunk/src/app/srs_app_hls.cpp b/trunk/src/app/srs_app_hls.cpp
index 8b36ab7f94..64a8396c0d 100644
--- a/trunk/src/app/srs_app_hls.cpp
+++ b/trunk/src/app/srs_app_hls.cpp
@@ -646,6 +646,11 @@ int SrsHlsMuxer::update_acodec(SrsCodecAudio ac)
     return current->muxer->update_acodec(ac);
 }
 
+bool SrsHlsMuxer::pure_audio()
+{
+    return current && current->muxer && current->muxer->video_codec() == SrsCodecVideoDisabled;
+}
+
 int SrsHlsMuxer::flush_audio(SrsTsCache* cache)
 {
     int ret = ERROR_SUCCESS;
@@ -1049,25 +1054,6 @@ int SrsHlsCache::write_audio(SrsAvcAacCodec* codec, SrsHlsMuxer* muxer, int64_t
         return ret;
     }
     
-    // flush if buffer exceed max size.
-    if (cache->audio->payload->length() > SRS_AUTO_HLS_AUDIO_CACHE_SIZE) {
-        if ((ret = muxer->flush_audio(cache)) != ERROR_SUCCESS) {
-            return ret;
-        }
-    }
-
-    // TODO: config it.
-    // in ms, audio delay to flush the audios.
-    int64_t audio_delay = SRS_CONF_DEFAULT_AAC_DELAY;
-    // flush if audio delay exceed
-    // cache->audio will be free in flush_audio
-    // so we must check whether it's null ptr.
-    if (cache->audio && pts - cache->audio->start_pts > audio_delay * 90) {
-        if ((ret = muxer->flush_audio(cache)) != ERROR_SUCCESS) {
-            return ret;
-        }
-    }
-    
     // reap when current source is pure audio.
     // it maybe changed when stream info changed,
     // for example, pure audio when start, audio/video when publishing,
@@ -1083,6 +1069,14 @@ int SrsHlsCache::write_audio(SrsAvcAacCodec* codec, SrsHlsMuxer* muxer, int64_t
         }
     }
     
+    // directly write the audio frame by frame to ts,
+    // it's ok for the hls overload, or maybe cause the audio corrupt,
+    // which introduced by aggregate the audios to a big one.
+    // @see https://github.com/simple-rtmp-server/srs/issues/512
+    if ((ret = muxer->flush_audio(cache)) != ERROR_SUCCESS) {
+        return ret;
+    }
+    
     return ret;
 }
     
@@ -1100,7 +1094,7 @@ int SrsHlsCache::write_video(SrsAvcAacCodec* codec, SrsHlsMuxer* muxer, int64_t
         // do reap ts if any of:
         //      a. wait keyframe and got keyframe.
         //      b. always reap when not wait keyframe.
-        if (!muxer->wait_keyframe()|| sample->frame_type == SrsCodecVideoAVCFrameKeyFrame) {
+        if (!muxer->wait_keyframe() || sample->frame_type == SrsCodecVideoAVCFrameKeyFrame) {
             // when wait keyframe, there must exists idr frame in sample.
             if (!sample->has_idr && muxer->wait_keyframe()) {
                 srs_warn("hls: ts starts without IDR, first nalu=%d, idr=%d", sample->first_nalu_type, sample->has_idr);
@@ -1110,9 +1104,6 @@ int SrsHlsCache::write_video(SrsAvcAacCodec* codec, SrsHlsMuxer* muxer, int64_t
             if ((ret = reap_segment("video", muxer, cache->video->dts)) != ERROR_SUCCESS) {
                 return ret;
             }
-            
-            // the video must be flushed, just return.
-            return ret;
         }
     }
     
diff --git a/trunk/src/app/srs_app_hls.hpp b/trunk/src/app/srs_app_hls.hpp
index 62abd660f9..a9269d69ea 100644
--- a/trunk/src/app/srs_app_hls.hpp
+++ b/trunk/src/app/srs_app_hls.hpp
@@ -309,6 +309,10 @@ class SrsHlsMuxer
     virtual bool is_segment_absolutely_overflow();
 public:
     virtual int update_acodec(SrsCodecAudio ac);
+    /**
+     * whether current hls muxer is pure audio mode.
+     */
+    virtual bool pure_audio();
     virtual int flush_audio(SrsTsCache* cache);
     virtual int flush_video(SrsTsCache* cache);
     /**
diff --git a/trunk/src/kernel/srs_kernel_codec.hpp b/trunk/src/kernel/srs_kernel_codec.hpp
index c47628276c..cc995dd046 100644
--- a/trunk/src/kernel/srs_kernel_codec.hpp
+++ b/trunk/src/kernel/srs_kernel_codec.hpp
@@ -246,12 +246,6 @@ extern int aac_sample_rates[];
 #define SRS_SRS_MAX_CODEC_SAMPLE 128
 #define SRS_AAC_SAMPLE_RATE_UNSET 15
 
-// in ms, for HLS aac flush the audio
-#define SRS_CONF_DEFAULT_AAC_DELAY 60
-
-// max PES packets size to flush the video.
-#define SRS_AUTO_HLS_AUDIO_CACHE_SIZE 128 * 1024
-
 /**
 * the FLV/RTMP supported audio sample size.
 * Size of each audio sample. This parameter only pertains to
diff --git a/trunk/src/kernel/srs_kernel_ts.cpp b/trunk/src/kernel/srs_kernel_ts.cpp
index 0ae7d58a52..864e05aa8d 100644
--- a/trunk/src/kernel/srs_kernel_ts.cpp
+++ b/trunk/src/kernel/srs_kernel_ts.cpp
@@ -2760,6 +2760,11 @@ void SrsTSMuxer::close()
     writer->close();
 }
 
+SrsCodecVideo SrsTSMuxer::video_codec()
+{
+    return vcodec;
+}
+
 SrsTsCache::SrsTsCache()
 {
     audio = NULL;
@@ -3134,20 +3139,9 @@ int SrsTsEncoder::write_audio(int64_t timestamp, char* data, int size)
         return ret;
     }
     
-    // flush if buffer exceed max size.
-    if (cache->audio->payload->length() > SRS_AUTO_HLS_AUDIO_CACHE_SIZE) {
-        return flush_video();
-    }
-
-    // TODO: config it.
-    // in ms, audio delay to flush the audios.
-    int64_t audio_delay = SRS_CONF_DEFAULT_AAC_DELAY;
-    // flush if audio delay exceed
-    if (dts - cache->audio->start_pts > audio_delay * 90) {
-        return flush_audio();
-    }
-
-    return ret;
+    // always flush audio frame by frame.
+    // @see https://github.com/simple-rtmp-server/srs/issues/512
+    return flush_audio();
 }
 
 int SrsTsEncoder::write_video(int64_t timestamp, char* data, int size)
diff --git a/trunk/src/kernel/srs_kernel_ts.hpp b/trunk/src/kernel/srs_kernel_ts.hpp
index e20778afc6..3003f6cf32 100644
--- a/trunk/src/kernel/srs_kernel_ts.hpp
+++ b/trunk/src/kernel/srs_kernel_ts.hpp
@@ -1586,6 +1586,11 @@ class SrsTSMuxer
     * close the writer.
     */
     virtual void close();
+public:
+    /**
+     * get the video codec of ts muxer.
+     */
+    virtual SrsCodecVideo video_codec();
 };
 
 /**