ML Suno Repair Lab v41.9 — audio restoration workstation

Source / Transport

Drop a WAV / FLAC / MP3 / M4A / OGG here

Browser decoding support depends on the installed codecs. WAV is recommended for restoration.

OPTIONAL STEM PRE-PROCESSING · BEFORE MASTER REPAIR For vocals/drums/bass/other work, separating first can improve role-aware repair in StemLab before the final master pass. This step is optional and can be skipped because neural separation can take several minutes.

optional · not chosen

Source→Optional Splitter→Optional StemLab→Master Repair→Listen→Validate→Export

Engine Output

Load audio to estimate the optional stem workload.

Recommended default: Auto chooses HTDemucs for 2/4 stems. Six stems require BS-RoFormer. No stem processing is started until you press Split first.

Analyze = diagnosis only · Smart Restore = analyze → recommendations → calibration → validation.

1 SOURCE load audio

2 ANALYZE measure defects

3 RESTORE build repair

4 LISTEN A/B · Delta · ABX

5 VALIDATE quiet · median · dense

6 EXPORT WAV / Δ

Session · load a source to begin

CLARITY

PUNCH

SPACE

TEXTURE

SAFETY

50 = baseline; below 50 reduces the macro, above 50 increases it. BASE captures the current rack as the new reference without resetting the DSP.

Repair Assistant macros: neutral baseline · Smart Restore or a strategy will position these controls from measured evidence.

5-KNOB REPAIR DESK Fast control surface: Source Repair · Clarity · Dynamics · Space · Safety. It drives the existing guarded DSP, not a second processing path.

SOURCE REPAIR CLARITY DYNAMICS SPACE SAFETY

EXPERT MACRO MATRIX Route each macro to up to three DSP/global targets with bipolar depth.

00:00.000 / 00:00.000

STEREO MASTER VU · RMS + PEAK RMS bar · instantaneous peak marker · red Peak/RMS hold per channel

PK −∞ · RMS −∞ dBFS

A 00:00.000 · B 00:00.000

ABX idle

A/B/C B is captured after Smart Restore · C is the current manual state

Focus Spectrum Visual style live

Lane Paint directly on the defect map. Shift = wider brush · Alt = pinpoint.

WebGL2 · holographic stereo vectorscope · persistence + correlation field

phase —

WebGL2 · luminous spectrum analyser · 20 Hz – 20 kHz · pre / post + centroid

peak —

WebGL2 · oscilloscope · waveform

corr —

WebGL2 · topographic defect map · luminous hotspots + editable timeline

run analysis

SOURCE · PROCESSED · defect map: BASS / VEIL / GLARE / SPACE / TEXTURE / PHASE / TRANSIENT . The smoothed map drives module-specific automation; it visualizes evidence and is not a provenance detector.

WebGL2 · adaptive 3D waterfall · Scientific / Studio / Spectacular depth profiles

WebGL2 · layered chain response · global curve + per-module spectral anatomy

TRANSIENT SKYLINE · energy + onset hotspots · click to seek

LOW-END PHASE FIELD · <150 Hz · coherence risk

Holographic multiband vectorscope · LOW amber / MID cyan / HIGH violet

Before / After / Delta spectrogram · split, delta or tri-view

Before / After technical scorecard

run validation

— Collapse

— Texture

— Phase

— Dynamics

— HF Stress

— Mono compatibility

— Crest

— Loudness Δ

Over-mastering guard: waiting for validation.

INPUT

OUTPUT

K-weighted loudness-match A/B Wet 100% Match +0.0 dB · GR 0.0 dB

Offline diagnosis

— median spectral centroid

— median Side / Mid

— median crest factor

— persistent 17–20 kHz lines

— level-dependent collapse index

— high-band tail / 20 kHz cliff

— persistent mid/high resonances

— gated K-loudness range estimate

— gated K-LUFS estimate

— 5–10 kHz fizz / sibilance ratio

— low-correlation section risk

— source sample / est. 4× intersample peak

— presence / low balance

— observed generator / provenance

— dense-section bass rise / air loss

— Side/Mid low · mid · high

— 5–16 kHz spectral flatness

— minimum-statistics noise floor

— inter-frame phase-swish index

— spectral persistence score

— harmonic trajectory confidence

— dense Δ sub · body · 2–5 · 5–10 · 10–16 kHz

— texture flatness quiet → dense

— repair-profile inference / hint

— dense Δ Side low · mid · high

— dense HF texture / glare stress

— legacy phase / bandwidth repair need

— spatial intent / topology

— measured drift evidence confidence

— self-reference / adaptive match target

Scene-aware processing 85%

Healthy-zone gate 10%

Voice-aware De-Ess / De-Harsh

Module-aware automation: bass, veil, glare, space, phase, texture and transient evidence drive only matching DSP modules. Below the Healthy-zone gate corrective gain/width modules are truly at 0%, eliminating residual over-processing in clean sections; only safety rumble filtering and limiting remain available.

Safety governor 80%

Max DSP delta −5.0 dB

Phase floor 0.02

Backs off the global repair mix if the DSP delta exceeds the chosen transparency ceiling or processed stereo correlation becomes unsafe.

Target lock + Delta Guard 90%

Low Δ budget −2.5 dB

Mid Δ budget −4.5 dB

High Δ budget −6.0 dB

Target Lock stops learned EQ near the calm-reference ratio. Delta Guard independently caps correction energy below 180 Hz, 180 Hz–8 kHz and above 8 kHz before the global governor.

Self-reference Learn none

Mark a clean source passage with A/B and click Learn Ref A–B , or let diagnosis pick the healthiest section with Auto Ref . Manual A–B remains authoritative. Frequencies above 16 kHz are never boosted by this module.

Multi-scene Reference Atlas 100%

Learns quiet/medium/dense anchors plus up to three spectrally distinct healthy scenes. In realtime the matcher chooses the closest healthy atlas nodes using both level and nine-band tonal shape, instead of forcing every chorus toward one verse-like target. Manual A–B overrides the atlas.

Mono Compatibility Guard −9.0 dB

Caps side-only correction energy relative to the dry Mid signal. It preserves intentional width while preventing repair layers that sound impressive in stereo but largely disappear or change balance after mono fold-down.

Run analysis to build a band-by-band repair plan.

Analysis hotspots will appear here.

Diagnosis: Load a source and run analysis. The stationary-line detector is an artifact detector, not a Suno watermark decoder.

Chain validation: not run. Validation renders a short representative excerpt through the current AudioWorklet chain and checks width, crest factor, peak level and phase safety.

Synthetic DSP benchmarks: not run. Tests bypass, tonal preservation, minimum-statistics noise cleanup, phase-swish repair, persistence/tonal tracking and de-click independently of the loaded song.

Render

Audacity · EBU R128 · EBU Radio

FINAL WAV · LOCKED · VALIDATE required

True-Peak FIR ×8 TP —

HQ offline spectral render · 2048-point STFT; realtime preview stays 1024-point. Ultra offline spectral render · 4096-point STFT (~11.7 Hz/bin at 48 kHz). Overrides HQ for validation/export; slower but better for very narrow stationary artifacts.

Source + rendered hashes, before/after metrics, curves, validation and full DSP state.

Export performs an offline render through the same AudioWorklet chain. Δ correction exports the actual current Wet/Governor result minus the latency-aligned source for DAW blending/automation. The source C2PA manifest is not copied into a repaired WAV because the derivative has new bytes and a new hash. The optional Suno Credentials test can check whether Suno's verifier can still recover provenance for the derivative.

ML Suno Repair Lab

Source / Transport

RESTORATION ASSISTANT

Before / After technical scorecard

Offline diagnosis

Render

Stem Splitter · dual neural engines

Splitter workflow

StemLab · source / mixer

Stem channels

Inter-stem analysis

StemLab render

Welcome