Re-Production Studio

Make your AI song sound unmistakably human. Starter runs Direct AI Vocal Humanization — the AI vocal stops reading as AI — and rebuilds the instrumental with real instruments, so the whole song passes AI detection. Pro goes further and adds your own voice — a private trained voice or a live sing-along. Follow the six steps below.

🏠 Home 📖 Tutorial 💸 Affiliates — Earn recurring income 🎻 PRO — Re-Perform with real instruments & human feel 100% Client-Side Processing Fully Automatic

🎚️ Your Workflow

Step 1 of 6 — Start.

1 Upload the AI Song

Start here. Drop the complete AI song (the full mix — AI lead vocal + any backing vocals + instrumental, as one MP3 or WAV) and pick one goal. That single choice shapes the whole studio — you are never asked again. We check it for copyright the moment you choose it, then separate the stems automatically.

What do you want to do?

Pick one to shape your workflow.
🎵
No file loaded
MP3 or WAV. Your song is checked for copyright the moment you choose it, then the stems are separated automatically — no extra clicks.

What happens next

  • Step 2 — the same song is separated into lead vocal, backing vocals & instrumental automatically. The lead vocal becomes the app’s “AI Vocal”.
  • Step 3 — Voice: we show the one face your goal picked — humanize the AI vocal, re-sing it in your trained voice, or sing along & record. No re-asking.
  • Step 4 — Rebuild: we rebuild the AI bed with real instruments so the whole song passes.
  • Step 5 — Master: the automatic pipeline builds one finished, mastered track, with the AI-likeness meter and the 3-stage fidelity gate.
Waiting for a song…

2 Stem Split the AI Song SERVER

We separate the song you dropped in Step 1 into a clean lead vocal, optional backing vocals, and an untouched instrumental — automatically, the moment your song is uploaded. The lead vocal becomes the app’s “AI Vocal” and the instrumental becomes the “Backing Track”, so the Master step can blend your voice under the AI vocal and remix one finished master. Already have stems? Upload them below for the best accuracy.

We separate your song automatically — use this only if you want to run it again.

Lead Vocal

Not split yet

Backing Vocals

Not split yet

Instrumental

Not split yet
🔒
You need an active plan to split stems. Choose Starter or Pro above to continue.

3 Humanize the Vocals

RP-3  |  Vocal bus  |  Auto-verify
TAKE 01 Starter

Humanize the AI vocal so it stops reading as AI — no singing required. This step is vocals only; you rebuild the instrumental in the next step.

Direct AI Vocal Humanization Chosen in Step 1 — change it there any time.
STARTER

This is the standard path and needs no recording. In the full flow it also runs automatically in the Master step — the controls below let you fine-tune it.

🧬 Direct AI Vocal Humanization STARTER

The standard workflow. Re-synthesizes the AI vocal so it stops reading as AI, then verifies the result against a real, published AI-music detector so you can see exactly how much it changed. It works directly on the AI vocal stem itself — even if you never record a real vocal. Runs on our servers.

Detector: checking… Engine: spectral de-fingerprint (near-identical)
EngineSpectral de-fingerprint (auto) — the setting that produced clean, human vocals.
AUTOLocked
Humanization strengthWinning value — the highest fidelity-gate score we measured.
35%Locked
Breath / airHeld at zero. The engine auto-cleans the source if a setting would fail the gate.
0%Locked
VerifyEvery render is scored against the real detector and the fidelity gate.
ONLocked
Advanced · reference vocal & manual scoring
No reference — uses the AI vocal's own tone
Upload a real human vocal to borrow its vocal-tract colour. Optional — the engine removes the AI fingerprint either way.

Engine, strength and breath are locked to the winning preset — there is nothing to tune. Use Score Current AI Vocal to measure the source before you run.

🎯 Vocal AI-Detection Likeness Meter lofcz/ai-music-detector · fakeprint (ISMIR 2025)
Before
—
After
—
Run the engine to measure how the detector scores your vocal before and after.

4 Re-Perform the Instrumental PRO (optional)

This is the step that makes the whole song pass. Your AI vocal is already humanized in Step 3 — now re-perform the AI instrumental with real instruments and human feel. We transcribe the bed to MIDI, humanize the performance (timing, velocity, groove, micro pitch drift), then re-render it through real-instrument models — so the bed keeps the same musical content but gets a brand-new waveform with none of the AI generator’s spectral fingerprint. The switch below is armed by default so the whole song passes.

Both Starter and Pro rebuild the instrumental so the whole song passes AI detection. Pro goes further and adds your own voice (trained voice or sing-along) to blend a real human voice into the song.

PRO
This is what makes the whole song pass. Tick it to re-perform the AI instrumental with real instruments and human feel — we transcribe it to MIDI, humanize the performance (timing, velocity, groove, micro pitch drift), then re-render it through real-instrument models. The bed keeps the same musical content but gets a brand-new waveform with none of the AI generator's spectral fingerprint. Off by default; turn it on to include it in this render.

Your choice is remembered: Re-perform includes the real-instrument bed in the next step’s render; Skip keeps the AI instrumental untouched. You can change this any time before you run Automatic Processing.

AVH Studios · Master Bus RP-5  |  −14 LUFS  |  −1.0 dBTP
TAKE 01 Starter
  • VocalsHumanized · gated
  • →
  • BedRebuilt · gated
  • →
  • SumBalanced
  • →
  • MasterClean · limit · score
Humanize backing vocals Same DSP on the AI backing so it doesn't carry the fingerprint into the mix.
Remove AI fingerprints Spectral, timing, dynamics, stereo, noise floor, metadata — all below audibility.
Master to streaming −14 LUFS integrated, −1.0 dBTP true peak. Export-ready without clipping.
Blend a real vocal −15 dB under the lead vocal
Match AI Sound AUTO · 85%
Match the mix to the original AI song Static spectral-envelope match — magnitude only, so it cannot re-introduce the generator fingerprint. Runs automatically before mastering.
85%
Source · stems & take

AI Song Stems from Step 2

Not split yet — run Step 2 first
Lead, backing and instrumental are reused from your Step-2 split.

Your Vocal from Step 3

No take yet — record in Step 3, or load a file here
A take recorded in Step 3 is reused here automatically.
Master out ● Standby
Vocal
Bed
Master
LUFS−14.0
True peak−1.0 dBTP
Final Mix Bay Enhance · EQ · audition · download
Vocal Enhancer → Rack EQ → Final Mix
Preview the Finished Master WYSIWYG · what you hear is what you download
0:00/0:00
–∞
Run Automatic Processing to load your finished master. This plays the exact file you download — same bytes, same mastering.
Auto-Balance Vocals → Instrumental
Run Automatic Processing first.
Vocal
—
Inst
—
Ratio
—
-7.4
🎚️ Vocal Enhancer PRO The full Xvox-style vocal chain — dynamics, tone, space, SFX, the Enhancer knob & Atlas reverb — now applied to the finished mix. Press ▶ Play on the monitor above and tune it live; Step 6 saves exactly what you hear. ▶
Rack Mixer · 4-Band EQ UNITY
0.0 dB
0.0 dB
0.0 dB
4-Band Rack EQ 90 Hz shelf · 350 Hz bell · 2.5 kHz bell · 8 kHz shelf
Preset Open-source curves · EasyEffects community presets (GPL-3.0)
301003001k3k10k20k
+120-12
0.0
0.0
0.0
0.0
Live — faders and EQ change the sound instantly. The bottom of each throw mutes that layer. Bake final mix to render the mastered download.

You need an active plan to run Automatic Processing. Choose Starter or Pro above to continue.
Live processing & pitch comparison

As soon as your Step-2 stems and Step-3 take are loaded, the humanization chain applies automatically — no buttons to press.

  • ○ Time-aligning real vocal to AI vocal
  • ○ Analyzing sibilance & breath dynamics
  • ○ Analyzing & gently correcting pitch
  • ○ Applying micro-timing drift & consonant emphasis
  • ○ Building EQ / duck / de-ess / saturation chain
  • ○ Ready to preview

Live pitch comparison

Blue = AI vocal pitch (target)  •  Orange = your real vocal pitch  •  Green dashed = after gentle auto-correction.

Blue = AI vocal  •  Orange = Real vocal, auto-aligned.

Studio controls (advanced — optional, Pro)

Everything here already has a sensible default applied automatically. Open this only if you want to fine-tune by hand (Pro feature).

Advanced settings (optional — everything already has sensible auto-applied defaults)
📖 How to use Advanced Settings (tutorial) — click to expand

Everything below already has a sensible default applied automatically. This reference explains what each control does if you want to fine-tune it by hand (Pro feature).

Timing

Fine-tune timing offset. The app auto-aligns your vocal, but if you want to nudge it earlier or later by hand, use the −100ms/−10ms/+10ms/+100ms buttons or type an exact millisecond value. The offset display shows what the automatic analysis originally detected.

Tone Shaping

Real Vocal Level (Duck). How far below the AI vocal your real vocal sits in volume, in dB. Recommended range is 12–18 dB below — enough to add texture without competing for attention.

High-Pass Filter. Cuts low-frequency rumble and handling noise from your recording.

Low-Mid Cut (Mud). A targeted dip around the low-mids (default 300 Hz) to stop your vocal from making the mix sound muddy when stacked underneath the AI vocal.

Low-Pass "Body" Filter. The key move in the whole chain — rolls off the higher frequencies of your real vocal so it acts as warm "body" and texture rather than a second, competing lead vocal.

Presence Tuck (optional). A narrow cut around 3 kHz if your real vocal is still poking through too much on top of the AI vocal.

Warmth Shelf (optional). A low-shelf boost around 200 Hz if the blended layer feels too thin.

Saturation / Warmth Drive. Soft, analog-style harmonic saturation. Higher values add more grit and warmth; 0% keeps it clean.

Realism Features

De-esser. Automatically detects and tames harsh "S"/"SH" sibilance frame-by-frame. Adjust the amount of reduction and target frequency, or disable it.

Breath Preservation. Naturally-recorded breaths are lifted slightly above the duck level so they remain audible — this preserves your real breaths, it never synthesizes fake ones.

Consonant Emphasis. Detects the sharp attack of consonants like P, B, T, and K and gives them a small boost, since real voices attack these harder than AI vocals do.

Micro-Timing Drift. Adds an extremely subtle, slowly-modulating delay so your vocal isn't perfectly locked to the beat at the sample level — mimicking natural human timing instability.

Pitch Correction ("Melodyne-Lite")

Pitch Correction Strength. How strongly your vocal's pitch is pulled toward the AI vocal's pitch. Defaults to a gentle 40% on purpose — full "snap to pitch" correction can make a real voice sound more robotic, not less.

Max Correction Range. Caps how many semitones any single note can be pulled, limiting artifacts on notes that are very far off.

Use pitch-corrected real vocal. Uncheck to instantly A/B your original, uncorrected recording against the corrected version.

Utility Buttons

Reset to Recommended Defaults. Returns every Advanced Setting to the app's tested default values in one click.

Re-run Automatic Processing. After changing any Advanced Setting, click this to reprocess your vocal with the new values.

Want the full walkthrough including the basic workflow? See the full tutorial page.

🎤 Backing Vocal Take OPTIONAL

By default, Automatic Processing humanizes the AI backing vocals with DSP (the toggle in the Master step). If you'd like to go further, you can record your own backing take here — it will be time-aligned, pitch-matched and layered under the AI backing vocals using the same “Nuclear Option” technique as the lead. This is for power users; the DSP pass alone is enough for most songs.

No backing take — optional
00:00
Fine-tune timing offset: ms Offset: not analyzed yet
Recommended: 12–18 dB below the AI vocal
Removes rumble / handling noise
Bell cut to prevent mud when stacked under the AI vocal
The key move — makes the real vocal act as body, not presence
Narrow cut ~3 kHz if the real vocal still pokes through
Low-shelf ~200 Hz if the layer feels too thin
Soft analog-style harmonic saturation — adds grit/warmth, 0% = clean
Auto-detects & tames sibilance (S/SH sounds), frame by frame
Lifts naturally-recorded breaths above the duck level. Preserves, does not synthesize, breaths.
Auto-detects P/B/T/K attack transients and boosts them slightly — real vocals attack consonants harder than AI vocals do
Very subtle slow delay modulation so the layer isn't perfectly time-locked — mimics natural human micro-timing instability
Deliberately kept gentle by default — see "Why pitch correction is kept subtle" below
Caps how far any single note can be pulled, to limit artifacts
Uncheck to A/B against the uncorrected recording

Final mix & fidelity gate

Tune the re-production, then A/B it against the original to confirm it still sounds like your song. This is exactly what gets exported.

00:00 / 00:00
🎚️ Faithful Re-Production
The original AI lead stays the lead; your real voice is blended as a subtle, pitch-matched layer underneath it — humanizing the sound without changing it.
Layer alignment —
A/B compare
Switch instantly to hear the difference.
AI Vocal
0 dB
C
Real Vocal
0 dB (on top of auto-duck)
C
Backing Track
0 dB
C
🎚️ Master Bus PRO Live mastering chain — EQ, saturation, glue, limiter & reference A/B. Click to open. ▶

🤖 AI Mastering

One click analyses your mix and applies a complete, pro mastering chain — EQ, resonance clean-up, analog tape, punch, glue and limiting — then loudness-matches it to your target. Everything stays editable below.

GentleBalancedAggressive
Balanced
45%
35%
50%

Pick an intensity and target, then hit AI MASTER.

Pick your genre and the master is tuned to that style's tonal balance and loudness — automatically.

A real mastering chain on the full mix, live: 4-band EQ → saturation → glue compression → stereo width → brick-wall limiter. Everything you tweak here is applied identically to the exported WAV — true WYSIWYG.

Preset:

4-Band EQ

Dynamics & Tone

Output & Limiter

Reference A/B

Load a commercial song to loudness-match and A/B against your mix. Its average spectrum becomes the tonal-balance target curve above.

No reference loaded
Peak-∞
RMS-∞
LUFS~-∞
Gain Red.0.0 dB

Tip: raise Drive and Glue for a warmer, denser, more "finished" master. Keep Peak below 0 dBFS and aim for LUFS~ −9 to −14 for streaming-ready loudness.

Step 6 · Save & Download

Congratulations — your song is complete

Give it a name, save it to My Songs, and download the finished master. The name you type here is the name on the file and in your library.

My Song.wav

This names your download and your saved song.

🔒
You need an active plan to save. Choose Starter or Pro above to continue.

Your master lands in My Songs the moment you save it — under the title above.