diff --git a/player/audio.c b/player/audio.c
index 3a2c60b8ab..2ce1669366 100644
--- a/player/audio.c
+++ b/player/audio.c
@@ -886,13 +886,18 @@ void fill_audio_out_buffers(struct MPContext *mpctx)
 
     assert(mpctx->audio_status >= STATUS_FILLING);
 
+    // We already have as much data as the audio device wants, and can start
+    // writing it any time.
+    if (mpctx->audio_status == STATUS_FILLING)
+        mpctx->audio_status = STATUS_READY;
+
     // Even if we're done decoding and syncing, let video start first - this is
     // required, because sending audio to the AO already starts playback.
-    if (mpctx->audio_status == STATUS_FILLING && mpctx->vo_chain &&
-        !mpctx->vo_chain->is_coverart && mpctx->video_status <= STATUS_READY)
-    {
-        mpctx->audio_status = STATUS_READY;
-        return;
+    if (mpctx->audio_status == STATUS_READY) {
+        if (mpctx->vo_chain && !mpctx->vo_chain->is_coverart &&
+            mpctx->video_status <= STATUS_READY)
+            return;
+        MP_VERBOSE(mpctx, "starting audio playback\n");
     }
 
     bool audio_eof = status == AD_EOF;