diff --git a/render_audio.py b/render_audio.py
new file mode 100755
index 0000000..9e2cddf
--- /dev/null
+++ b/render_audio.py
@@ -0,0 +1,72 @@
+#!/usr/bin/python
+import os
+import time
+import sys
+import json
+
+import mlt
+from PyQt5 import QtWidgets
+
+# Avoid segfault in webvfx
+app = QtWidgets.QApplication(sys.argv)
+
+#mlt.mlt_log_set_level(40)  # verbose
+mlt.Factory.init()
+
+
+tractor = mlt.Tractor()
+tractor.mark_in = -1
+tractor.mark_out = -1
+
+multitrack = tractor.multitrack()
+
+source = sys.argv[1]
+target = source.replace('.json', '.audio.xml')
+
+with open(source) as fd:
+    data = json.load(fd)
+
+music = mlt.Playlist()
+vocals = mlt.Playlist()
+
+fps = 60
+profile = mlt.Profile("atsc_1080p_%d" % fps)
+
+def add_clip(playlist, file_, duration):
+    in_ = 0
+    if not isinstance(file_, str):
+        file_ = file_.encode('utf-8')
+    clip = mlt.Producer(profile, file_)
+    clip.set_in_and_out(in_, in_+duration-1)
+    playlist.append(clip)
+
+def add_blank(playlist, length):
+    playlist.blank(length)
+
+
+for clip in data['music']:
+    frames = int(clip['duration'] * fps)
+    if clip.get('blank'):
+        add_blank(music, frames)
+    else:
+        add_clip(music, clip['path'], frames)
+
+for clip in data['vocals']:
+    frames = int(clip['duration'] * fps)
+    if clip.get('blank'):
+        add_blank(vocals, frames)
+    else:
+        add_clip(vocals, clip['path'], frames)
+
+multitrack.connect(music, 0)
+multitrack.connect(vocals, 1)
+composite = mlt.Transition(profile, "mix")
+composite.set("combine", 1)
+
+tractor.plant_transition(composite)
+
+consumer = 'xml'
+consumer = mlt.Consumer(profile, consumer, target)
+consumer.connect(tractor)
+#consumer.set("real_time", -2)
+consumer.start()