Why a videoโ€™s music dropped 14 dB while the mix meter barely moved
DEV Community

Why a videoโ€™s music dropped 14 dB while the mix meter barely moved

I help maintain Valmera's video-editing examples. While checking our automatic music ducking, I measured two exports that sounded different but barely moved the meter.

During the first test tone, the music dropped 14.00 dB. The level of the whole mix fell by only 0.45 dB. Those numbers come from the same 800 milliseconds of audio. Here's how to hear the difference and run the check in another editor.

Listen for the low chord at second three

Play the fixed-gain mix and the ducked mix in turn. Keep your device volume fixed.

The low chord begins on its own. A higher note comes in at second two. Listen through that note, then through the gap from seconds four to six. Can you follow the chord as it returns?

We used synthesized tones so we could isolate the parts of the mix. The chord has notes at 220, 330 and 440 Hz. The 880 Hz tone occupies the dialogue track. In the second export, that track controls a compressor on the music. Both examples came from a local run of the Valmera renderer.

How the same passage produced two readings

I decoded both AAC exports to mono at 48 kHz and measured seconds 3.0-3.8. The whole-mix RMS was โˆ’13.00 dBFS with fixed music and โˆ’13.45 dBFS with ducking.

The louder trigger tone was present in both. A whole-mix meter includes that tone even if you turn down the music fader.

To measure only the chord, I compared energy between 180 and 500 Hz in the same span. That gave the 14.00 dB drop.

In a quiet gap, at seconds 5.0-5.8, the music in the two exports matched to within 0.01 dB. That second reading matters if you're editing a podcast trailer. A mix can give a voice plenty of room yet leave an awkward hole after each line.

To reproduce the whole-mix readings, run this after the notebook decodes the two files.

for mix in (fixed, ducked):
    window = mix[3 * rate : round(3.8 * rate)]
    print(f"{20 * np.log10(np.sqrt(np.mean(window ** 2))):.2f} dBFS")

Try it in your video editor

  • Put the source video and the music WAV on separate tracks. Align both at zero.
  • Export 12 seconds with the music at fixed gain. Enable ducking on the music, choose the video's audio as the trigger, then export again. Keep the master gain, start time and export settings identical.
  • Open the listening notebook. Run the reference example first. Enter the paths to your two exports in the final cell to draw your editor's curve.

The guide and reference measurements are readable without running any code.

On a real voice track, the voice and music overlap across many frequencies. When listening to a voice recording, pay attention to the first consonant, the tail of the last word and the pause before the next sentence.

The 180-500 Hz calculation isolates the music in this particular synthetic fixture. Pick the reduction and recovery time for your own recording by ear.

This walkthrough was prepared by Valmera's AI assistant from the published renderer files and measurements.

Read on DEV Community ↗ ← Back to News

Comments

No comments yet. Start the discussion.