Proof, not promises
Hear What KROMA Does
Every pair below is a real recording run through the same pipeline your file would take. Press play, then switch between the original and the result mid-playback — the difference is the product.
Crowd & Noise Removal
The flagship: the recording's environment becomes independent faders. Pull the crowd or the wind down — the performance itself is never filtered.
Interactive A/B Demo
Fancam with Screaming Fans, Rescued
The crowd layer is separated and pulled down while the vocal stays untouched — switch mid-playback to hear the singer emerge.
Open in Studio →Demo: “Let It In” by Josh Woodward — CC BY 4.0; crowd — live audience recording
Interactive A/B Demo
Concert Recording — Roar Turned into Ambience
Total silence would sound dead; the crowd is kept at a low level so the recording keeps its live energy while the music leads.
Open in Studio →Demo: “Epic Pop Dance” by TeknoAXE — CC BY 4.0; crowd — live audience recording
Interactive A/B Demo
Voice Memo — Background Music Removed
The music under the narration sits on its own stems — muting the instruments leaves the voice intact, without the hollow artifacts of a broadband noise gate.
Open in Studio →Demo: Tom Sawyer — LibriVox narration (public domain); “Wallpaper” by Kevin MacLeod — CC BY 3.0
Mastering & Polish
The mastering chain on real tracks — loudness, dynamics and tone shaped toward platform targets.
Interactive A/B Demo
Indie Pop, Polished for Release
The All-Round Polish chain: EQ, glue compression and limiting bring a rough mix to release loudness without squashing the dynamics.
Open in Studio →Demo: “Cherubs” by Josh Woodward — CC BY 4.0
Interactive A/B Demo
Electronic Track → Club Master
Midnight Club tightens the low end and brings the master to club loudness — the kick hits harder without clipping the limiter.
Open in Studio →Demo: “Epic Pop Dance” by TeknoAXE — CC BY 4.0
Interactive A/B Demo
Classical Recording, Cleaned & Restored
Pristine Clear removes the haze from an older recording and restores definition — listen to the decay of each phrase.
Open in Studio →Demo: Pachelbel's Canon in D — public domain
Stem Separation & Mixing
One stereo file, independent faders. Mute a sound entirely or just pull it down — the rest of the mix stays untouched.
Interactive A/B Demo
Full Rock Mix → Studio Acapella
The lead vocal is isolated from drums, bass and guitars — breaths and phrasing stay intact, no metallic residue.
Open in Studio →Demo: “Swansong” by Josh Woodward — CC BY 4.0
Interactive A/B Demo
Drums Muted, Everything Else Intact
Muting the drum stem removes kick, snare and hi-hats cleanly while the vocal and chord accompaniment stay untouched.
Open in Studio →Demo: “Swansong” by Josh Woodward — CC BY 4.0
Interactive A/B Demo
Produced Track → Intimate Unplugged
The Intimate Unplugged mode rebalances the stems toward a close, acoustic feel — like the band moved from stage to living room.
Open in Studio →Demo: “Swansong” by Josh Woodward — CC BY 4.0
Voice & Instrument Conversion
A neural conversion model re-synthesizes a melody as a different voice or instrument — phrasing preserved, timbre replaced.
Interactive A/B Demo
Violin Lead → Converted Voice
Pitch tracking follows every phrase of the lead violin line while the conversion model re-synthesizes it as a different persona — the backing stays untouched.
Open in Studio →Demo: Pachelbel's Canon in D — public domain
Interactive A/B Demo
Sung Vocal → Violin
The vocal melody is re-synthesized as a violin — vibrato and phrasing carry over, the lyrics become instrumental lines.
Open in Studio →Demo: “Swansong” by Josh Woodward — CC BY 4.0
Interactive A/B Demo
Sung Vocal → Saxophone
Same melody, different horn — the conversion keeps the singer's phrasing and dynamics on a saxophone timbre.
Open in Studio →Demo: “Swansong” by Josh Woodward — CC BY 4.0
Space & Atmosphere
Reshape where a recording sounds like it was made — from dry studio to concert hall.
Interactive A/B Demo
Dry Strings → Symphonic Hall
Air Symphonic convolves the stems with a real hall impulse response — width and tail bloom without washing out the lead.
Open in Studio →Demo: Pachelbel's Canon in D — public domain
Under the hood
First-Tier Separation, Measured
KROMA's vocal and acapella paths run a Roformer-family model (BS / Mel-Band Roformer) on GPU — the architecture family that leads public separation benchmarks such as the MVSEP and MUSDB-18 leaderboards. We verified it on controlled mixes: a known dry vocal plus a known backing track, recombined, then separated.
+11.4 dB
vocal SDR vs Demucs, dense backing
+3.4 dB
vocal SDR vs Demucs, vocal-led mix
~1×
realtime on GPU (T4)
12
interactive A/B demos on this page
Method: scale-invariant signal-to-distortion ratio between the separated vocal and the known dry vocal (VocalSet), on 0 dB and −3 dB vocal/backing mixes. Same comparison available in our studio: Demucs remains the fallback engine for non-vocal stems.
Free to try — no install, works in the browser.