Guides

Why AI songs stop at 16 kHz, and why an EQ boost cannot fix it

Updated 2026-09-06

Put a Suno or Udio export on a spectrum analyzer and look at the right edge. Somewhere between 15 and 17 kHz the level falls off a cliff, often by 60 dB or more. A commercial release keeps going to 20 kHz. That missing band is the "air" of a record, and its absence is the first thing a trained ear notices.

Why it is there

Generators work on a compressed representation of audio, and most of them were trained at a sample rate or a codec setting that discards the top band. The model never learned what a cymbal sounds like above 16 kHz, so it never produces it. Newer versions move the cliff up a little, but it is still a cliff.

Why a boost fails

An EQ can only make louder what is already there. Above the cutoff there is nothing, so a high shelf lifts the noise floor and the last few hundred hertz below the cliff. The result is hiss and a sharper edge, not air.

What a rebuild does

Bandwidth extension takes the harmonics that do exist between 8 and 16 kHz and generates their upper partials, at the level the music implies, then blends them in below the cliff so there is no seam. Done well, cymbals get their shimmer back and vocals get breath, and nothing sounds added because it follows the music note for note.

How to check it worked

• The spectrum should slope gently down to 20 kHz, not stop or spike.
• Switch between before and after at the same loudness. Louder sounds better on its own, so match levels before you judge.
• Listen to a snare hit in isolation. Rebuilt air rides on the hit and decays with it. Boosted hiss does not.

Burnish measures the cutoff on your file, rebuilds only above what it finds, and shows the cutoff frequency and the rebuild on the report.