KROMA
← Journal
4 min read

Removing Crowd Noise from Fancams — What Works, What Nobody Can Do

Every fancam has the same problem: the recording captures the stage, but also fifty people screaming, cheering and singing along two feet from the phone mic. Removing that layer is one of the most-requested audio fixes — and also one of the most misrepresented.

What actually works

Audience noise — cheering, applause, screaming, chatter — has a recognisable acoustic signature. A model trained specifically on concert crowds can estimate the crowd layer, and a frame-level classifier can tell which moments of the recording actually contain audience sound versus pure performance. Subtract only where crowd is present, and the result is a dramatically cleaner mix.

The harder part is protecting the singer. Naive crowd models leak the lead vocal into the crowd estimate — subtract it and you hollow out the singer. The fix is to isolate the vocal stem first and estimate crowd from what's left, so the singer is never in the input the crowd model sees.

The honest limit: fans singing along

A fan singing along is a pitched human voice on the same melody — it lives inside the same vocal stem as the singer. Separating one voice from another singing the same line is called speaker-attributed separation, and on phone-mic concert audio it remains unsolved: pitch can't tell them apart (same notes), and voice fingerprints trained on speech degrade badly on sung, reverberant, PA-blurred vocals.

This is not a KROMA limitation — it applies to every tool on the market, from free UVR models to commercial services. Cheers and screams come out; a sing-along fan may leave residue. We'd rather state that plainly than promise a result nobody can deliver.

How we handle it

The crowd engine runs an iterative estimate: separate the vocal, estimate crowd from the residual, remove it with a spectral mask that leaves vocal-dominated cells untouched, then estimate once more on what remains — measured at roughly a third of crowd energy removed with the singer intact. The vocal is never subtracted; if the model can't tell crowd from music, it stops rather than muting your mix.