SRL — Complete Product Guide & Technical Specification

01 / PURPOSE

What SRL is for

ML Suno Repair Lab is a browser-based workstation for diagnosing, reducing and checking audible defects in existing music recordings—especially the tonal, textural, stereo and dynamic artifacts encountered in generated music.

It brings Master Repair, neural Stem Splitter and StemLab into one workspace. You can repair a stereo master directly, or separate it into parts, work on those parts, return the mix to Master Repair, and prepare a WAV for delivery. Audio-to-MIDI transcription provides a separate route from those parts to editable musical notes.

The guiding idea is measured, selective intervention: find what changes between healthy and problematic passages, apply the relevant correction, listen to what has been changed, and check the result before export. A loud chorus is not automatically a defect; a narrow mix is not automatically something to widen.

For music creators

Refine an existing song, reduce distracting shimmer or harshness, rebalance a crowded passage, compare versions and prepare a listening copy.

For producers and editors

Use role-aware stems, masking analysis, deterministic humanization, correction-only WAVs, MIDI files, reusable presets and project states.

For technically minded listeners

Inspect diagnostic timelines, stereo fields, spectral evidence, per-module activity, before/after measurements and reproducible reports.

Implementation: Application scope and workflows · DSP catalogue and initial parameter values

02 / FROM SOURCE TO DELIVERY

One recording. Two repair routes.

The stem stage is optional. The final Master workflow stays the same whether the input is an imported file or a mix returned from StemLab.

Load source→Analyze→Smart Repair / manual DSP→Listen & validate→Normalize & export

Load source→Optional neural separation→StemLab→Send mix to Master Repair→Final Master workflow

Stage	What happens	Your control
Source	Import an existing recording. The application creates processing state and derivative outputs rather than overwriting the original file.	Browse or drag and drop; use a supported direct-file picker or a launch URL.
Prepare	Choose direct Master Repair or optional stem separation.	Skip stems, or choose a separator and 2/4/6 output groups.
Repair	Analyze, apply recommendations, calibrate strength or adjust the rack manually.	Conservative, Balanced and Aggressive strategies; macros; individual module controls.
Listen	Compare the Master input, current DSP result and correction signal.	A/B, Delta, A/B/C, blind ABX, looping, hotspots and final-WAV audition.
Validate	Render representative scenes and check measured changes for safety and overshoot.	Read the results; revise and rerun after changing the audio chain.
Deliver	Choose normalization, offline spectral quality and WAV encoding.	Audition the encoded WAV, then save the same prepared asset.

The guided tutorial follows this workflow. It offers the optional stem branch, waits for required operations, brings the relevant controls into view, explains the defect timeline and Delta, includes validation, and presents a normalization destination selector before final export. The tutorial normalization controls and the export panel share the same settings.

Implementation: Application scope and workflows · Final loudness engine, presets and controls

03 / CAPABILITY MAP

The complete workspace at a glance

SRL combines restoration, component-level work, musical transcription and delivery checks. The detailed sections below separate what each function does from the conditions it needs.

Measure and locate

Offline diagnosis; nine-lane defect timeline; density-linked tonal drift; stereo topology; resonances; transient and noise-texture evidence; healthy reference selection.

Repair and protect

18 reorderable DSP modules; 108 module parameters; scene automation; learned matching; correction-energy budgets; mono and phase guards; peak safety.

Separate and remix

Two neural separator integrations; 2/4/6 output modes; up to 16 imported or separated stems; per-stem repair, humanization, spectral ducking and residual glue.

Recover editable notes

Seven transcription modes; tonal consensus and dedicated drum analysis; four quality settings; note cleanup, velocity reconstruction, pitch bends and multitrack MIDI export.

Compare and finish

Named playback provenance; A/B, A/B/C, Delta and ABX; exact encoded-WAV pre-audition; LUFS/RMS normalization; ten destination choices; three WAV sample formats.

Work and document

Four layouts; three usage levels; five-knob desk; 16 themes; 27 interface languages; shortcuts, MIDI control, gamepad, projects, batch processing, reports and diagnostic tools.

Implementation: DSP catalogue and initial parameter values · Tonal/drum transcription and MIDI serialization · Supported launch switches and startup execution · Platform services, projects, autosave, gamepad and corpus

04 / DIAGNOSIS

Find the problem before changing the sound

Analyze source is a measurement operation, not a repair command. It produces a diagnosis, a band-by-band plan and a timeline that can drive subsequent processing.

Evidence family	Measurements and indicators	How it is used
Level and dynamics	RMS and peak information, crest factor, K-weighted integrated loudness and range estimates, transient density and level-dependent collapse.	Distinguish quiet, median and dense passages; detect lost punch or excess intervention.
Tonal balance	Spectral centroid, low/presence balance, low-band rise and signed changes in sub, 500 Hz–2 kHz, 2–5 kHz, 5–10 kHz and 10–16 kHz regions.	Separate bass takeover, body masking, high-frequency veil and high-frequency glare.
Stereo and mono	Side/Mid energy, low/mid/high Side/Mid ratios, L/R correlation, low-correlation sections, mono-fold behavior and inferred spatial intent.	Protect an intentionally narrow image; respond to measured width collapse or over-expansion.
Narrow artifacts	Persistent 17–20 kHz lines and persistent mid/high resonances.	Supply detected frequencies to conditional line and resonance processing.
Texture and motion	5–16 kHz spectral flatness, minimum-statistics noise-floor estimates, inter-frame phase-swish, spectral persistence and harmonic-trajectory confidence.	Provide evidence for texture, noise, shimmer and spectral correction.
Context and references	Healthy passages, quiet/medium/dense anchors, profile hints, historical acoustic similarity and evidence confidence.	Guide recommendations without letting a generation label overrule the measured recording.

The principal narrow-line scan uses 16,384-point FFT windows sampled across the file. Density-band analysis uses 4,096-point windows; the spectral-intelligence analysis uses 2,048-point windows. Probe counts depend on the performance profile. These are targeted, sampled diagnostics—not continuous laboratory measurements of every possible defect.

Actual SRL defect-map interface after analyzing an illustrative synthetic diagnostic signal. These scores are not a music-restoration benchmark.

Nine editable defect lanes

Bass, Veil, Glare, Space, Texture, Shimmer, Click, Phase and Transient. The map shows estimated intensity over time. You can paint a lane, soften its curve or reset edits; Shift widens the brush and Alt narrows it. The smoothed map supplies module-specific automation. Hotspots provide navigation to passages worth checking.

The map has its own drawing path, independent of live playback and WebGL startup. It redraws when its view or size changes. Missing analysis, unusable measurements and genuinely zero scores are different states. The displayed map refers to the analyzed source; it is not automatically a new diagnosis of the final rendered WAV.

Implementation: Measurement, reference and analysis logic · Live/offline spectral profiles, graph and transport

05 / ADAPTIVE CONTROL

Repair where it is needed. Back off where it is not.

Several layers govern the amount and placement of correction. They complement the module settings rather than replacing listening.

Control	What it changes	Important distinction
Scene-aware processing	Maps the current time to bass, veil, glare, space, texture, shimmer, click, phase and transient evidence.	Matching modules receive matching evidence; the entire chain is not driven by one generic severity meter.
Scene strength	Controls how strongly evidence modulates the repair.	A global automation setting, separate from the final Wet mix.
Healthy-zone gate	Suppresses eligible corrections below a chosen evidence threshold.	Safety-oriented rumble filtering and limiting remain available; this is not a universal bit-identical bypass claim.
Voice-aware De-Ess / De-Harsh	Uses vocal-sensitive context to moderate corrective upper-band processing.	A protection mechanism, not a standalone vocal separator.
Safety governor	Reduces global correction when the DSP difference becomes excessive or stereo correlation becomes unsafe.	Uses the selected correction ceiling, governor amount and phase floor.
Target Lock	Restrains learned equalization around the calm-reference relationship.	Helps avoid overshooting a learned target rather than maximizing brightness.
Three-band Delta Guard	Caps correction energy in low, middle and high regions.	Bands are below 180 Hz, 180 Hz–8 kHz and above 8 kHz; budgets are independent.
Mono Compatibility Guard	Limits side-only correction energy relative to the dry Mid signal.	Designed to keep impressive stereo changes from disappearing or shifting balance on mono fold-down.
Spatial-intent anchoring	Constrains mid/high Side energy relative to the source topology.	An intentionally narrow or mono-core source is not automatically widened.
Global Wet	Blends the actual processed contribution with the latency-aligned input.	The optimizer is allowed to choose zero Wet when its measurements show no useful benefit.

Self-reference and the Reference Atlas

Mark a healthy A–B passage and use Learn Ref A–B to create an authoritative nine-band target. Auto Ref can select healthy material automatically. The multi-scene Atlas combines quiet, medium and dense anchors with up to three additional spectrally distinct healthy scenes. Matching considers both level and nine-band tonal shape instead of forcing every chorus toward a single verse-like target.

Manual A–B reference takes priority over the automatic Atlas. You can add an A–B passage as a good scene, load an external reference, and choose which reference categories to apply: Tone, Stereo, Dynamics and Texture. External matching is category-limited, not whole-master cloning. The adaptive Match EQ never boosts frequencies above 16 kHz.

Implementation: Live/offline spectral profiles, graph and transport · Measurement, reference and analysis logic · DSP catalogue and initial parameter values

06 / COMPLETE DSP CATALOGUE

18 modules. Every exposed module parameter.

The catalogue follows the initial rack order. You can reorder, bypass, solo, inspect, collapse and audition modules. The 108 parameter rows below are extracted from the v41.9 module definitions.

Actual English SRL rack: assistant macros, snapshots, module search, full module names and persistent playback identity.

DSP / SUB

Sub Control

On in initial rack

A rumble high-pass stage, a low shelf and envelope-triggered bass attenuation share the task of controlling low-end energy. It is aimed at headroom and balance, not at replacing bass notes.

When to consider it: When dense sections become bass-heavy or headroom is being consumed by 20–140 Hz.

All 4 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Low shelf	-5 to 2	dB	0.1	-1.2
Dynamic tame	0 to 100	%	1	25
Bass trigger	-38 to -18	dB	0.5	-28
Detector cutoff	80 to 180	Hz	1	125

DSP / DECLICK

De-Click / De-Crackle

Off in initial rack

Uses one-sample look-ahead to compare a suspicious sample with its neighbors and interpolate isolated discontinuities. Crackle detail lowers the threshold for smaller repeated impulses. Attack protection reduces intervention on musical transients.

When to consider it: Use only when the diagnosis and listening test justify it.

All 4 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Repair amount	0 to 100	%	1	72
Sensitivity	0 to 100	%	1	60
Crackle detail	0 to 100	%	1	30
Transient protect	0 to 100	%	1	88

DSP / DEMUD

De-Mud

On in initial rack

Uses dynamic low-mid attenuation rather than a permanently deep low-mid cut. Amount and threshold determine how much congested material is restrained.

When to consider it: When low-mid congestion masks vocals, guitars or percussion.

All 2 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 100	%	1	20
Sensitivity	-42 to -16	dB	0.5	-28

DSP / DENSITY

Defect-Keyed Signed Spectral Matrix

Off in initial rack

A signed, five-region correction system. It learns calm-to-dense drift and can key the intervention to the defect currently present. Positive and negative regional corrections distinguish a veil from excess glare; the context threshold is not a command to repair every dense passage.

When to consider it: When calm and dense sections show a repeatable signed tonal drift.

All 10 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Matrix amount	0 to 100	%	1	65
Density context	-36 to -12	dB	0.5	-21
Defect key	0 to 100	%	1	0
Key tolerance	0 to 6	dB	0.1	1.5
Max sub tame	0 to 6	dB	0.1	2.2
500 Hz–2 kHz Δ	-4 to 5	dB	0.1	0.7
2–5 kHz Δ	-4 to 5	dB	0.1	1
5–10 kHz Δ	-4 to 5	dB	0.1	1
10–16 kHz Δ	-4 to 4	dB	0.1	1.2
Release	80 to 900	ms	10	320

DSP / MATCH

Self-Reference Adaptive Match EQ

Off in initial rack

Nine-band matching to a learned internal target or an active reference bank. Amount, separate boost/cut limits, release, scene linking and energy neutralization constrain the response. Manual A–B reference and automatic Atlas selection are different targeting modes.

When to consider it: When healthy parts of the same song provide a trustworthy tonal reference.

All 6 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Match amount	0 to 100	%	1	42
Max boost	0 to 4	dB	0.1	2.2
Max cut	0 to 6	dB	0.1	3.2
Release	120 to 1200	ms	10	520
Scene link	0 to 100	%	1	88
Energy neutralize	0 to 100	%	1	90

Additional state: nine-value reference curve, reference metadata and a bank of up to six nine-band reference scenes.

DSP / PRESENCE

Presence Recovery

On in initial rack

Broad intelligibility and articulation adjustments around 1.8 kHz and 4 kHz. The two signed gain controls support modest recovery or reduction rather than indiscriminate high-frequency boosting.

When to consider it: For broad intelligibility recovery after bass/low-mid masking.

All 2 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Body / intelligibility	-2 to 4	dB	0.1	1
Clarity	-2 to 4	dB	0.1	1.2

DSP / DEHARSH

De-Harsh

On in initial rack

Dynamic restraint in the approximately 2.3–7.5 kHz upper-mid region. It addresses intrusive metallic or spitty energy and can be moderated by the voice-aware protection layer.

When to consider it: For metallic upper mids that become intrusive only on some passages.

All 2 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 100	%	1	12
Sensitivity	-44 to -18	dB	0.5	-29

DSP / DEESS

Fizz / De-Ess

Off in initial rack

A fast split-band process for roughly 5–11 kHz fizz and sibilance, with an adjustable detector focus. It is distinct from the slower broad upper-mid De-Harsh stage.

When to consider it: For synthetic sibilance, fizz or overactive 5–11 kHz vocal energy.

All 3 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 100	%	1	18
Trigger	-46 to -18	dB	0.5	-30
Focus	4500 to 9000	Hz	50	6500

DSP / SPECTRAL

Spectral Intelligence Engine

Off in initial rack

Linked-stereo short-time Fourier processing combines narrow-peak repair, texture attenuation, a minimum-statistics noise tracker, phase-swish reconstruction, persistence modeling and harmonic-trajectory protection. The sub-processes have independent amount, frequency and protection controls. Live FFT size is performance-profile dependent; offline FFT size follows the export-quality setting.

When to consider it: For narrow stationary peaks or noise-like texture that conventional EQ cannot target cleanly.

All 26 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Peak repair	0 to 100	%	1	38
Peak excess	2 to 12	dB	0.1	5
Peak max cut	0 to 12	dB	0.1	4.5
Neighborhood	3 to 14	bins	1	7
Transient protect	0 to 100	%	1	78
Peak low	1800 to 9000	Hz	100	3200
Peak high	8000 to 19000	Hz	100	16000
Texture cleanup	0 to 100	%	1	0
Texture max cut	0 to 5	dB	0.1	1.8
Texture flatness gate	0.45 to 0.9	unitless	0.01	0.68
Texture density link	0 to 100	%	1	55
Texture density gate	-36 to -12	dB	0.5	-22
Texture low	3500 to 12000	Hz	100	5500
Texture high	9000 to 18000	Hz	100	15000
Min-stat noise cleanup	0 to 100	%	1	0
Noise SNR margin	2 to 14	dB	0.2	8
Noise max cut	0 to 8	dB	0.1	3.5
Noise low	1200 to 10000	Hz	100	3500
Noise high	7000 to 19000	Hz	100	16000
Phase-swish repair	0 to 100	%	1	0
Swish threshold	0.15 to 1.8	rad	0.01	0.55
Swish max blend	0 to 80	%	1	42
Swish low	1200 to 10000	Hz	100	3500
Swish high	7000 to 19000	Hz	100	15500
Persistence model	0 to 100	%	1	72
Tonal trajectory protect	0 to 100	%	1	88

DSP / SHIMMER

Shimmer Suppressor

Off in initial rack

Tracks fast-versus-slow envelope instability across four high-frequency bands. The channels share the correction decision, with sensitivity, release, maximum attenuation and transient protection controlling how moving shimmer is reduced.

When to consider it: Use only when the diagnosis and listening test justify it.

All 7 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 100	%	1	42
Sensitivity	0 to 100	%	1	58
Low edge	3000 to 9000	Hz	100	4200
High edge	9000 to 18500	Hz	100	16000
Max cut	0 to 9	dB	0.1	4
Transient protect	0 to 100	%	1	88
Release	25 to 500	ms	5	160

DSP / RESONANCE

Adaptive De-Resonance

Off in initial rack

Uses the resonance frequencies supplied by analysis. Dynamic detection and a bounded depth reduce an identified resonance when it becomes intrusive; the display also exposes the learned frequency list.

When to consider it: For persistent narrow resonances identified by offline analysis.

All 4 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 100	%	1	32
Trigger	-48 to -18	dB	0.5	-31
Max depth	0 to 9	dB	0.1	4
Detector Q	3 to 24	Q	0.5	9

Additional state: analysis-selected resonance frequencies.

DSP / AIR

Safe Air / Exciter

On in initial rack

Combines a restrained high shelf and excitation with a gentle upper-frequency guard. This is optional finishing. It must not be confused with reliable recovery of information missing from the recording.

When to consider it: For restrained high-frequency finishing; avoid using it to compensate for genuinely missing information.

All 4 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Air shelf	0 to 3	dB	0.1	0.5
Exciter	0 to 60	%	1	10
Guard corner	15000 to 19500	Hz	50	17000
HF guard	0 to 12	dB	0.1	2.5

DSP / STEREO

Bidirectional Density M/S

On in initial rack

Separates static low/mid/high width from signed density-linked mid/high changes. Adaptive collapse response, correlation limits, source-intent anchoring and Side ceilings restrain inappropriate widening or narrowing.

When to consider it: For measured density-dependent width collapse or over-expansion, not simply because a mix is narrow.

All 13 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Bass Side	0 to 100	%	1	92
Base mid width	0 to 80	%	1	8
Base high width	0 to 90	%	1	10
Dense mid Δ	-50 to 60	%	1	0
Dense high Δ	-60 to 70	%	1	0
Density trigger	-36 to -12	dB	0.5	-22
Collapse adaptive	0 to 100	%	1	55
Low crossover	80 to 260	Hz	1	140
High crossover	1200 to 6000	Hz	25	3200
Correlation floor	0 to 70	%	1	18
Intent anchor	0 to 100	%	1	0
Mid Side ceiling	-30 to 0	dB	0.5	0
High Side ceiling	-30 to 0	dB	0.5	0

DSP / PHASE

Legacy Phase Coherence Guard

Off in initial rack

Contains low-band Side energy when measured L/R coherence becomes unsafe. Its frequency-selective focus, correlation floor, maximum Side attenuation and release are intended to avoid globally narrowing healthy material.

When to consider it: For early-generation low-band anti-correlation/phase instability.

All 5 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 100	%	1	55
Low focus	80 to 500	Hz	5	220
Low correlation floor	-0.2 to 0.8	unitless	0.01	0.12
Max low-Side cut	0 to 12	dB	0.1	6
Release	60 to 900	ms	10	320

DSP / HARMONIC

Legacy Harmonic Reconstructor

Off in initial rack

Resynthesizes upper-band harmonics from upper-mid material when an older source has a substantial density-linked bandwidth loss. This is deliberately reconstructive processing: the generated detail is plausible, not recovered original information.

When to consider it: Only when older generations genuinely lose upper-band information under density.

All 7 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 100	%	1	28
Density trigger	-36 to -12	dB	0.5	-21
Source low	1800 to 5500	Hz	50	2600
Source high	4500 to 9000	Hz	50	6800
Target low	5500 to 11000	Hz	50	7000
Target high	11000 to 17500	Hz	50	15500
Harmonic drive	1 to 6	x	0.1	2.6

DSP / TRANSIENT

Transient Restorer

On in initial rack

Uses fast/slow envelope contrast to emphasize attacks and recover some crest contrast. Amount is intentionally bounded; the validation stage checks for excessive transient expansion.

When to consider it: When crest factor collapses and attacks become flattened.

All 2 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Amount	0 to 50	%	1	12
Sensitivity	0 to 100	%	1	60

DSP / NOTCH

Conditional Line Guard

Off in initial rack

Uses analysis-detected stationary line frequencies and neighboring-energy context. A line is attenuated only when its excess rises above the chosen threshold; depth, Q and release remain adjustable.

When to consider it: For persistent stationary line artifacts that rise above neighboring spectral energy.

All 4 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Max depth	0 to 9	dB	0.1	3
Detector Q	18 to 90	Q	1	52
Excess trigger	2 to 12	dB	0.1	5
Release	60 to 800	ms	10	260

Additional state: analysis-selected stationary-line frequencies.

DSP / LIMITER

Look-Ahead Peak Safety

On in initial rack

A stereo-linked look-ahead limiter with four-times cubic intersample-peak estimation, plus latency accounting for dry monitoring and offline output. It is a peak-protection stage, distinct from the constant-gain final normalization engine.

When to consider it: Final peak protection. It is a safety stage, not a loudness-maximization tool.

All 3 parameters · ranges, increments and initial values

Parameter	Range	Unit	Step	Initial value
Ceiling	-3 to -0.1	dB	0.1	-1
Look-ahead	0 to 5	ms	0.1	2
Release	20 to 250	ms	1	90

Implementation: DSP catalogue and initial parameter values · Signal-processing engine and spectral implementation · Live/offline spectral profiles, graph and transport

07 / RECOMMENDATIONS & VALIDATION

Smart Repair, without hiding the decisions

The Smart Repair button orchestrates analysis, recommendations, strength calibration and validation. The separate commands remain available when you want to control each stage.

Diagnosis→Recommendations→Candidate calibration→Validation→Automatic comparison snapshot

Recommendations use the measured source, any profile hint, learned tonal and stereo context, and the available references. Conservative, Balanced and Aggressive strategies change how strongly SRL follows the measured repair target. Legacy v2/v3 modes can deliberately relax strict-conservation limits for more audible reconstruction.

Multi-objective calibration

The calibration routine evaluates a finite candidate set, including a null treatment, several Wet strengths and the current configuration. Candidates vary Wet, scene strength, Target Lock and governor amount. The routine renders each candidate across selected scenes and compares estimated benefit against safety flags, alteration cost, level drift, mono loss and texture penalties. It selects from a Pareto-filtered set with a weighted utility rule. This is not an exhaustive search of all 108 DSP parameters or a guarantee of a global optimum.

Representative-scene validation

With analysis available, SRL selects quiet, median and dense scenes, nominally 10 seconds each and bounded by the source. An enabled loop of at least four seconds can become the validation target, capped at 24 seconds. Without analysis, a representative excerpt is used. Scene rendering runs the embedded DSP in a reusable Worker, with progress based on processed samples.

Checks include	Examples of what is flagged
Stereo and mono	Unsafe correlation, a substantial Side increase, worsening mono-fold loss, width changes that conflict with source intent.
Dynamics and headroom	Crest loss, excessive transient expansion and very low peak margin.
Tone and texture	Amplifying the extreme upper band, over-lifting air, over-thinning sub bass, changing a quiet scene too strongly, or making dense texture noisier.
Current-state validity	Changes to the source or audio-chain fingerprint invalidate earlier validation; an unrelated old result does not unlock the current final export.

Chain Order Advisor

The advisor compares the existing rack topology with an analysis-aware recommended order, shows the proposed path and its reasoning, and lets you apply the order with optional recalibration. It preserves a distinction between an ordering suggestion and a validated result. Reordering nonlinear, dynamic or learned processors can change what later stages receive.

Long repair work exposes phase/candidate/scene progress and cancellation. A scene Worker that stops reporting advancement for 90 seconds produces an explicit stalled-worker error. On success, Smart Repair captures the automatic state for A/B/C comparison; manual changes then remain separately comparable.

Implementation: Smart Repair, validation and multi-objective calibration · Snapshots, presets and parameter assistance · Live/offline spectral profiles, graph and transport

08 / LISTENING & SIGNAL IDENTITY

Always know which signal you are hearing

The persistent bottom transport names the player, the source provenance and the processing stage. A is not necessarily the untouched file originally imported.

Actual persistent transport in its stopped “B · Current DSP” state, identifying the Master input and the pre-normalization processing stage.

Signal / player	What you hear	What is not implied
A · Master input	The current input to Master Repair, without Master DSP or final normalization. It may be an imported file or a mix returned from StemLab.	“A” does not always mean the initial untouched recording.
B · Current DSP	The current Master input through the current live DSP, Wet, selected solo and applicable monitoring settings.	It is before final normalization and is not the most recently exported WAV.
Δ · Corrections only	The processed correction contribution relative to the latency-aligned input.	It is not a full mix or a separated instrument stem.
Final WAV	The playback buffer decoded from the WAV bytes prepared for export, including selected normalization and encoding.	It is not routed through the live Master DSP for a second pass.
StemLab mix	The StemLab mix with its current mixing/processing state, before a subsequent Master pass.	Sending the mix to Master Repair is a separate operation.
Individual audio previews	A separated or imported stem, or another identified component preview.	The bottom bar disables controls that do not apply to that player.
MIDI preview	A synthesized audition of detected notes, individually or as a combined MIDI mix.	It is a note-checking reference, not a timbre reconstruction of the original instruments.
Blind ABX · X	An unknown comparison signal during a blind trial.	The identity indicator deliberately does not reveal whether X is A or B.

A/B and Wet

Switch between the current Master input and repair. The main listening path has its own K-weighted loudness-matching option; the final-WAV reader has a separate source-only comparison option, off by default.

A/B/C comparison

A is the source, B is the automatic state captured after Smart Repair, and C is the current manual state. This version comparison is separate from snapshot slots also named A–D.

Delta listening

Audition the full correction, a selected module’s contribution or a defect-focused passage. The Legacy Extreme Delta option adds +12 dB for monitoring only; it does not raise the exported master.

Blind ABX

Listen to known A and B, hear X, submit Guess A or Guess B, and follow trial/correct counts. ABX checks detectability; it does not decide which sound is artistically better.

Loop and hotspot work

Set A and B, toggle looping, jump to A, navigate previous/next hotspots and use the guided listening challenge to examine the most relevant defect families.

Explicit player switching

Listen to final WAV switches intentionally to the delivery reader. Return to Master changes context without automatically starting playback. Position controls remain horizontal sliders.

Implementation: Live/offline spectral profiles, graph and transport · Playback owner and processing-stage identity · Shared encoded-WAV preview/export asset · Long-track planning and source-separation engines

09 / ANALYZERS

Visual context, not just moving meters

The analyzer workspace offers Overview, Stereo, Repair and Deep analysis views so that the screen can show the evidence relevant to the current task.

Display	Purpose and interaction
Stereo vectorscope	Shows stereo geometry and correlation context, including persistence-oriented presentation.
Spectrum analyzer	Source, processed, overlaid or difference spectrum; a nominal 20 Hz–20 kHz display with peak/centroid context.
Oscilloscope	Time-domain waveform display for monitoring behavior.
Defect map	Estimated repair-family intensity over time, with editable automation and hotspot information.
Spectral waterfall	A history-oriented spectral display with Scientific, Studio and Spectacular presentation choices, including a 3D-style mode.
Chain response	An approximate visual response of the global chain and its component anatomy. It is not an exact transfer-function measurement of nonlinear or time-varying DSP.
Transient skyline	Energy/onset overview with click-to-seek navigation.
Low-end phase field	A focused view of low-frequency coherence risk below approximately 150 Hz.
Multiband vectorscope	Separate low, mid and high stereo fields.
Before/after/Delta spectrogram	Split, Delta or tri-view comparison layouts.
Technical scorecard	Validation summaries for collapse, texture, phase, dynamics, high-frequency stress, mono compatibility, crest and loudness change.
Master and module meters	Stereo RMS and instantaneous peaks, per-channel hold/reset indicators, input/output levels, gain reduction, module activity and module warning badges.
Reference and signature displays	The Reference Atlas constellation and historical/local acoustic-signature overlays support separate reference and inspection tasks.

Freeze stops live visual motion. Focus controls and double-click focusing enlarge a chosen analyzer. Source/processed/difference choices change what is drawn. Performance profiles adjust analyzer sizes and refresh work. WebGL2 is used for the richer analyzer surfaces; the defect timeline retains an independent drawing path.

Implementation: Signal-processing engine and spectral implementation · Measurement, reference and analysis logic · Interaction layers, navigation and preferences

10 / NEURAL STEM SPLITTER

Separate first when the problem needs stem-level work

SRL integrates two neural separation paths. The requested output count determines which engine can be used; a frequency split is never presented as neural source separation.

Engine	Outputs	Execution and requirements
HTDemucs	Four internal stems: Vocals, Drums, Bass and Other. Two-stem mode groups the non-vocal components into Instrumental.	The standard Auto choice for 2/4 stems. ONNX Runtime WASM, worker-backed execution, on-demand model loading; the build labels its model at approximately 172 MB.
BS-RoFormer-SW	Six underlying components, including Guitar and Piano; the selected 2/4/6 grouping determines retained outputs.	The required path for six stems. A heavier ONNX/WebGPU-oriented pipeline; the build labels its FP16 model at approximately 336 MB. Commercial-safe defaults block its remote checkpoint download pending rights clearance.

You can reuse the current Master source, load another source for the splitter, or supply a compatible local .onnx model for the selected engine. Local models must match the expected architecture and input/output contract; loading an arbitrary ONNX file is not supported. Model downloads use pinned revisions and SHA-256 integrity checks. Cache clearing and engine-state controls are exposed.

Long-track processing and memory

Separation operates at 44.1 kHz stereo. The planning layer uses 7.8-second HTDemucs segments and 4-second BS-RoFormer segments, with a 75% step / 25% overlap. Outputs are assembled over the complete requested duration. The interface estimates the memory needed for the source, model, working data and retained stems before the run. Long-track support is segmented, not unlimited: decoded full-length inputs and outputs still consume memory.

The splitter reports model/runtime readiness, duration, estimated memory, mode, active engine, progress and results. It checks output count, channel structure, sample validity and coverage before publishing stems. Source channels and transferable inference input data are given private ownership to avoid sharing mutable buffers with playback.

Handoff and cancellation

Automatically send results to StemLab is enabled by default. It remains a user choice, and manual transfer is available. A single primary progress track avoids duplicate animated representations of the same inference. Playback is locked during the separation operation; cancellation and the final result state determine when controls are released.

Implementation: Long-track planning and source-separation engines · Live/offline spectral profiles, graph and transport

11 / STEMLAB

Mix, repair and humanize the components

StemLab is a component-level mixer placed before final Master Repair. It accepts separated results or independently imported stems, with up to 16 simultaneous stem channels.

Roles include Vocal, Drums, Bass, Pad, Lead, FX and Other. Initial role detection uses filenames and measured features; you can change the role manually. Role-dependent defaults determine a starting point for width, light repair and humanization. The stem stage is deliberately lighter than the full Master DSP rack.

Per-stem control	Range	Purpose
Gain	−24 to +12 dB	Set the stem level.
Pan	−100 to +100	Position the stem across the stereo field.
Width	0–200%	Control the stem’s stereo width.
Repair	0–100%	Set role-aware light correction.
Humanize	0–100%	Set the overall motion contribution.
Duck	0–100%	Set key-related attenuation.
Gain ride	0–1	Weight the smoothed level movement.
Pitch wander	0–1	Weight the micro-delay/pitch-motion component.
Width breath	0–1	Weight the slowly varying width component.
Air ride	0–1	Weight the upper-band movement.

Each channel also has mute, solo, role selection, removal, a waveform view, metering, an audio preview, a MIDI action and per-channel export. Audio/MIDI preview transports use explicit time positions. Global controls cover Master gain, residual/glue amount, humanization amount, duck key and role-aware spectral ducking.

Analyze stems

Measure role context, spectrum, transient density and phase; build an inter-stem masking matrix; compare the unprocessed stem sum with the loaded Master and inspect residual energy.

Smart Stem Mix

Apply role-aware gain staging, conservative masking-derived ducking and bass-centering decisions. It does not run the entire Master Smart Repair workflow separately on every stem.

Deterministic humanization

Smoothed gain, micro-delay/pitch, width and air movement is seeded from stable stem information. Preview and offline stem rendering use the same automation logic, rather than generating a different random performance on each pass.

Role-aware spectral duck

Choose Off, Vocal or Drums as the key. Vocal-key ducking focuses mainly on approximately 0.7–6 kHz; drum-key ducking focuses on low/mid energy instead of indiscriminately attenuating the full spectrum.

Residual / glue

Build the difference between the reference Master and the raw stem sum, audition it, blend it back or export it. This can preserve material not reconstructed by the stem sum; it can also bring unwanted content back.

Return to Master

Export individual treated stems, all stem WAVs, a treated mix or the residual. Send mix → Master Repair renders a new Master input, identified as a StemLab mix, ready for the final chain and validation.

Implementation: StemLab mixing, import and channel controls · Signal-processing engine and spectral implementation · Live/offline spectral profiles, graph and transport

12 / AUDIO → MIDI

From separated audio to editable musical events

The default SRL path combines browser tonal transcription with a dedicated drum transcriber. Optional MuScriptor integrations offer alternate or hybrid paths; their required models are not bundled.

Order	Transcription mode	Meaning
1	SRL Stem Consensus · browser	Default selection. Uses the SRL tonal-consensus / cleanup path and its dedicated drum analysis.
2	Hybrid SRL + MuScriptor Web · ONNX/WebGPU	Combines the SRL path with a user-supplied MuScriptor browser bundle.
3	Hybrid SRL + MuScriptor Server	Combines SRL with the configured local HTTP MuScriptor service; the server hybrid path has an SRL fallback when the service is unavailable.
4	Hybrid SRL + MuScriptor Native · WebGPU	Adds the native Safetensors/WebGPU path to the hybrid workflow.
5	MuScriptor Native · WebGPU	Direct native WebGPU transcription using compatible supplied weights and configuration.
6	MuScriptor Web · ONNX/WebGPU	Browser ONNX bundle path without the SRL consensus as the selected primary engine.
7	MuScriptor · local server	Transcription through an external local HTTP service.

Choose which audio is transcribed

Auto · cleaned, no humanize avoids transcribing decorative motion from the mixer. The other choices are the Treated StemLab stem and the Original separated/imported stem. This choice is separate from the transcription engine.

Tonal analysis and four quality settings

The SRL tonal path loads Spotify Basic Pitch on demand and applies role-aware post-processing. Fast uses one decoder view; High uses three; Max uses five; Ultra uses eleven plus session reconstruction. These counts describe decoding views, not eleven independent neural models.

Quality	Tonal window	Overlap
Fast	30 seconds	2 seconds
High	30 seconds	2 seconds
Max	24 seconds	3 seconds
Ultra	22 seconds	4 seconds

The available reconstruction and cleanup mechanisms include phase-safe stereo downmixing, stereo rescue when needed, onset refinement, multiband onset evidence, recovery of repeated articulations, adjacent-note stitching, role-specific pitch/duration/polyphony constraints, octave-ghost suppression, harmonic verification, key-aware ghost filtering, hybrid audio/note tempo estimation and audio-derived note velocities.

Timing, expression and drums

Timing can preserve the performance, apply Smart groove · soft 1/16, or quantize to 1/16, 1/8 or 1/4. Tempo can be entered or estimated. Tonal sensitivity and drum sensitivity each offer Conservative, Balanced and Sensitive settings. Independent options cover pitch bends, role-aware cleanup, key-aware filtering, velocity reconstruction, polyphonic-safe bends, cross-stem reconstruction and multi-hit drum detection.

The SRL drum path combines multi-resolution onset and spectral classification rather than passing drums through the tonal decoder. It outputs percussion events on MIDI channel 10. Tonal bends are smoothed/recentered and expressive overlapping notes can be allocated to separate MIDI channels. Combined export distributes non-percussion channels across the session; when expression cannot be isolated without collisions, bend preservation is limited rather than promised for unlimited polyphony.

Exports and audition

Transcribe and export a stem independently; export all MIDI as separate files; or build one combined multitrack Standard MIDI File. Combined export uses shared musical context and timing, includes role-based instrument assignments, and keeps drums on channel 10. Clear MIDI results invalidates stored note results. Per-stem MIDI audition and a synchronized MIDI mix preview help check note timing and content, with stop and horizontal seek controls.

Optional model paths in detail

Path	Configuration and implementation
Web / ONNX	Bundle URL or local bundle files, model loading, precision and execution choices, cache clearing, status and progress. Precision choices: Auto, FP16, INT8 and FP32. Execution choices: Auto, WebGPU and WASM proxy Worker. The build’s Auto policy targets FP16 on WebGPU and INT8 on WASM; an appropriate model bundle is still required.
Web model storage	SHA-256-checked IndexedDB caching; worker-backed mel preprocessing and WASM inference. FP32 is the compatibility path described by the build.
Native / Safetensors	Weights URL/file and model configuration, optional in-memory Hugging Face token, load/unload controls, checkpoint-specific STFT and mel filterbank, dedicated WebGPU Worker, optional FP16 GPU matrices when shader-f16 is available, F32 activations/KV cache.
Native fallback	An explicit option allows the ONNX browser path when native WebGPU cannot run. Compatible model assets are still needed; the fallback does not manufacture a missing bundle.
Server	Configurable local HTTP endpoint and a test action. The service runs outside the browser HTML and must be installed/configured separately.

Implementation: Tonal/drum transcription and MIDI serialization · StemLab mixing, import and channel controls

13 / EXPORT & NORMALIZATION

Listen to the file you are actually going to save

The final preview and processed Master export share one encoded-WAV asset. Normalization is part of preparation, not a separate level change applied after you listen.

Offline Master DSP→Optional peak safety→Optional normalization→WAV encoding→Exact pre-audition / save

SRL renders the Master, applies selected finishing options, encodes the WAV and constructs the preview buffer from those encoded samples. PCM quantization and dithering therefore occur before the final audition. Saving the processed Master uses those same prepared bytes while the settings remain current, rather than silently rendering a different version.

The final preview uses a separate reader at unity gain, outside the live DSP chain. “Loudness-match Source for fair A/B” is off by default. Turning it on changes only the source-comparison gain; it does not alter Final WAV playback. Changing the source, DSP, Wet, render quality, normalization, ceiling or bit depth invalidates an older prepared asset. A bit-depth-only change can reuse the normalized floating-point render and re-encode it.

WAV encoding and correction outputs

Output	Description
Processed Master WAV	16-bit PCM, 24-bit PCM or 32-bit floating point. Integer formats include dithering; floating output preserves finite values outside ±1 before any selected peak constraint. The Master render is stereo.
Full correction Δ WAV	The actual current correction, relative to the latency-aligned source, for inspection or controlled DAW blending. Final loudness normalization does not change this Delta.
Spectral / Stereo / Dynamics Δ	Grouped correction exports with only the relevant module families enabled. These are separate renderings, not a guaranteed additive decomposition of the nonlinear full chain.
Stem outputs	Individual treated stems, all stem WAVs, the treated StemLab mix and residual/glue output use the StemLab export path.
Reports and projects	JSON diagnosis/QA/corpus/preset files, scientific HTML reports, and .srl processing projects are separate from audio deliverables.

Normalization: on by default, TikTok selected

The initial final-level setting enables linked-channel LUFS normalization with the TikTok preparation preset, a target of −14 LUFS and an estimated peak ceiling of −1 dBTP. Saved projects and explicit user choices can change that starting state. The same destination choice is available in the tutorial before export.

Destination	Suggested target	Peak ceiling
Balanced listening	−16 LUFS	−1 dBTP
YouTube	−14 LUFS	−1 dBTP
Facebook	−14 LUFS	−1 dBTP
Instagram	−14 LUFS	−1 dBTP
TikTok · initial default	−14 LUFS	−1 dBTP
TV Europe · EBU R128 label	−23 LUFS	−1 dBTP
Radio · EBU programme label	−23 LUFS	−1 dBTP
CD preparation	−16 LUFS	−1 dBTP
Vinyl preparation	−18 LUFS	−3 dBTP
Custom	User-selected	User-selected

How final level is calculated

The LUFS mode uses K weighting, 400 ms blocks at 100 ms hops, an absolute −70 LUFS gate and a relative gate 10 LU below the ungated-above-threshold average. Clips shorter than 400 ms are treated as estimates. RMS mode includes silence. The engine measures the complete repaired buffer and applies constant gain; it does not add a compressor or limiter as part of normalization.

Linked processing preserves L/R balance. Independent-channel normalization is available and can change that balance. Mono/dual-mono compensation is configurable. The estimated peak ceiling takes priority over the loudness target: a peak-constrained result can remain quieter than the chosen LUFS target. The engine leaves a small rounding margin and reports before/after loudness, RMS, peak, gain and constraint status.

Final peak estimation uses eight-phase, 18-tap windowed-sinc interpolation. An additional True-Peak FIR ×8 export safety option uses this estimator and reduces gain when needed. This is separate from the rack’s look-ahead limiter and its four-times cubic estimate.

Implementation: Shared encoded-WAV preview/export asset · Final loudness engine, presets and controls · Live/offline spectral profiles, graph and transport

14 / REPEATABLE WORK

Batch processing, album balance and saved sessions

Use a queue for repeat work, and use snapshots and projects to keep processing decisions recoverable. These are processing-state tools, not a full multitrack DAW session format.

Batch and album tools

Add multiple audio files, remove items, analyze the queue and choose a strategy per row. Batch analysis summarizes level, centroid and width. The batch renderer uses the current shared chain with strategy-dependent Wet values—55% Conservative, 78% Balanced or 100% Aggressive. It is not an independent full Smart Repair optimization for every song.

Album consistency learns an average target from the analyzed queue. Optional corrections are restrained: gain adjustment is bounded to ±3 dB, high-shelf tonal adjustment to ±1.5 dB and Side-energy adjustment to ±2 dB. Final peak safety and final normalization can follow. Per-track normalization can reduce intentional inter-song level differences; disable it when those differences are part of the album design.

Snapshots, undo and preset exchange

Four snapshot slots, A, B, C and D, let you store and recall complete processing states. An A↔B morph interpolates compatible settings between the first two snapshots. The history timeline stores up to 40 states, with undo, redo and direct history-node recall. Presets can be saved/imported as JSON; migration logic supplies missing newer controls when older states are loaded.

.srl projects and autosave

Projects retain DSP settings, references, automation, snapshots/history and related session data with source fingerprints. They do not embed a complete multitrack audio session: audio must be relinked separately. Loading a project requires fresh analysis/validation of the current source before relying on a final-export gate.

Direct open/save actions use the browser file-system picker when available, with ordinary picker/download fallbacks where implemented. Crash-safe autosave stores session state in IndexedDB and can cache original source bytes up to the configured 96 MiB ceiling. Restore and Clear autosave are exposed. Browser storage can be refused, evicted or cleared; keep independent audio and project backups.

Implementation: Batch strategy and album adjustments · Snapshots, presets and parameter assistance · Platform services, projects, autosave, gamepad and corpus

15 / PRESETS & MACROS

Fast controls with a path to the full rack

Start with a simple control surface, a strategy or a factory preset, then open the underlying modules when a specific correction needs attention.

Repair Assistant macros

Clarity, Punch, Space, Texture and Safety route to the existing DSP/global controls. 50 is the current baseline; lower values reduce that macro contribution and higher values increase it.

Rebase macros

BASE captures the current rack as the new macro reference without resetting the DSP. Smart Repair and strategy selection can place the macro controls from measured evidence.

Five-knob repair desk

Source Repair, Clarity, Dynamics, Space and Safety provide a smaller control surface over the existing chain. This is not a second audio-processing engine.

Expert macro matrix

Each macro can route to up to three DSP/global targets with positive or negative depth. Mapping targets include presence, transient/sub controls, stereo width, spectral texture/noise, shimmer and protection budgets.

All 16 factory presets

01Conservative restorationconservative

02v2 / XXL legacy reconstructionv2legacy

03v3 / engine-I density reconstructionv3

04v3.5 / adaptive legacy repairv35

05v4 / narrow-core preservationv4

06v5 Cover / crow conservative repairv5cover

07v5 / carp adaptive repairv5carp

08v5.5 / chirp-fenix density repairfenix55

09Legacy collapse repairoldCollapse

10Chirp-hawk high-band guardhawkGuard

11v6-Pro + MaxMode density texture stabilizerhawkProMax

12v6-mini density repairgooseMini

13v6-mini MaxMode density + width repairgooseMax

14Adaptive full repair v29adaptive

15Spectral artifact repairspectralSafe

16Neutral / all effects flatneutral

Profile hints include Auto, generic/unknown and the historical v2 through v6-family labels recorded in the application. They serve as contextual priors. Historical source labels and the presets named after them are not a current product catalogue, verified model identification or an authorship verdict.

For supported legacy-source cases, Authentic, Reconstruction and Rescue offer increasing intervention. Reconstruction and Rescue intentionally relax some transparency constraints. The separate Extreme Delta audition makes the correction easier to hear without changing its export gain.

Implementation: DSP catalogue and initial parameter values · Snapshots, presets and parameter assistance · Interaction layers, navigation and preferences

16 / ERGONOMICS

A workspace that can reveal more as you need it

The interface combines a stage-based workflow with an expandable workstation. Its main controls are organized around source, repair, listening, validation and export rather than requiring every panel to stay visible.

Dimension	Available choices / behavior
Workspace layouts	Studio, Analyzer (wide), Console (compact) and Classic Rack. These are the runtime labels; launch IDs are studio, wide, compact and classic.
Desk	Advanced Studio / full workstation, or Simple Repair / five-knob desk.
Usage level	Novice, Advanced and Expert. Progressive disclosure hides specialized controls at lower levels; expert controls remain available without changing the underlying module catalogue.
Control style	Knobs, Hybrid (knobs + sliders) and Sliders. Temporal navigation stays a horizontal slider even when parameter controls use knobs.
Density	Comfortable, Compact and Dense. Smaller screens use larger touch targets and stage-focused navigation.
Navigation	Clickable sticky workflow stages, Master / Splitter / StemLab tabs, preferences drawer, collapsible panels, focused analyzer views and a persistent listening transport.
Rack management	Module search, enabled-only filtering, clear filter, collapse/expand, dedicated reorder grip, up/down actions, bypass, solo, Delta/audition and secondary ⋯ actions.
Context and focus	Selected-module Inspector directly below the chain, readable module names, activity/warning indicators and pulsing block focus. Focus is visual context, not audio gain reduction.
Guidance	Context bubbles, inline explanations or help off; searchable Help Center; guided tutorial with spotlight/arrow; hotspot listening challenges and command palette.
Keyboard accessibility	Focus-visible styling, native controls where applicable, preferences-drawer focus handling, Escape dismissal and shortcut guards while editing text or interacting with controls.
Motion	Reduced-motion styling keeps state readable without depending on animated pulse effects.
Startup feedback	PowerShell-inspired startup terminal, initialization stages and progress; a separate launch console documents and parses app options.

16 visual themes

Carbon MintDeep OceanCobalt StudioViolet NightRose QuartzAmber ConsoleEmerald RoomTeal PrecisionArctic GlassCopper TapeCrimson StageRoyal BlueMatrix GreenImpulse Tracker / DemosceneCRT AnalyzerDark Minimal

27 interface languages

Startup can select the interface language from launch options, saved settings or the browser/OS language. Arabic and Hebrew use right-to-left presentation while audio axes, numeric fields, meters and code-oriented content retain appropriate left-to-right handling.

enEnglishfrFrenchdeGermanesSpanishptPortugueseitItaliandaDanishfiFinnishisIcelandicnoNorwegiansvSwedishnlDutchruRussianzhChinesejaJapanesekoKoreanarArabicheHebrewlaAncient LatinplPolishcsCzechelGreektrTurkishhuHungarianroRomanianukUkrainianhiHindi

This document is the English product guide. The application includes the language set above; this is not a claim that every historical translated string has received independent linguistic certification.

Implementation: Interaction layers, navigation and preferences · Supported launch switches and startup execution · Snapshots, presets and parameter assistance

17 / HANDS-ON CONTROL

Keyboard, hardware MIDI and gamepad

These shortcuts describe the workstation, not this presentation page. Shortcuts are context-sensitive and are generally suppressed while typing or using dialogs.

Key / gesture	Workstation action
Space	Play / pause the Master transport.
B	Switch Master A/B when available.
D	Toggle correction Delta.
L	Toggle the audition loop.
F	Freeze / unfreeze analyzer motion.
[ / ]	Previous / next hotspot.
1 / 2 / 3 / 4	Recall snapshot A / B / C / D; an empty slot can capture a state.
Shift + click snapshot	Store the current state in that snapshot slot.
Ctrl/Cmd + Z	Undo; Shift + Ctrl/Cmd + Z performs redo.
Ctrl/Cmd + Y	Redo.
Ctrl/Cmd + K	Open the command palette; arrows, Enter and Escape navigate/run/close it.
A / R / V	Analyze / Smart Repair / Validate when available.
E / Shift + E	Processed export / correction export when their controls are available.
H / T	Help Center / guided tutorial.
X	Audition unknown X during an active ABX session.
Escape	Dismiss the applicable palette, tutorial, help or control dialog.
Knob drag / wheel / arrow keys	Adjust a parameter; Shift gives finer and Alt ultra-fine adjustment where implemented.
Double-click parameter knob	Open numeric entry in the enhanced rack controls.
Shift / Alt while painting map	Wider / pinpoint defect-map brush.

Web MIDI control

Connect an available hardware MIDI input, choose a target, arm Learn next CC, then move a hardware knob or fader. Mappings can address Wet, the five macros, the selected DSP parameter, A/B, Delta and Play/Pause. Mappings can be cleared and are stored locally. This controls SRL parameters; it is separate from audio-to-MIDI transcription.

Gamepad

Optional gamepad control maps rising button presses to playback, A/B, Delta, challenge advance and hotspot navigation. The sticks adjust Wet or the selected parameter, and the triggers adjust Safety. Device recognition, permissions and button availability depend on the browser/device; the interface shows the detected controller state.

Implementation: Platform services, projects, autosave, gamepad and corpus · Interaction layers, navigation and preferences

18 / ENGINEERING SPECIFICATION

What is happening inside the audio path

SRL uses embedded JavaScript DSP for restoration, browser audio graphs for monitoring/rendering, Workers for scene processing, and separate optional neural runtimes for separation and transcription.

Subsystem	Implementation characteristics
Master graph	AudioContext → source → repair-chain AudioWorklet → monitoring gain → analyzer/output. The worklet is configured for stereo output. A dry monitoring/analyzer path is maintained for comparison.
Memory and arithmetic	Audio samples and many FFT/ring buffers use Float32Array / AudioBuffer storage; analysis and final-level routines also use Float64Array accumulators. The WAV sample format is an export choice, not a global arithmetic-precision setting.
Spectral process	Radix-2 FFT-based linked-stereo STFT. Square-root Hann windows, hop size N/4, overlap-add output. Supported process sizes: 512, 1,024, 2,048 and 4,096 samples.
Live spectral size	Laptop: 512; Normal: 1,024; Ultra preview: 2,048. This is separate from the larger display-analyzer FFT.
Offline spectral size	Standard: 1,024; HQ: 2,048; Ultra: 4,096. Ultra overrides HQ. At 48 kHz, 4,096 samples gives approximately 11.72 Hz bin spacing; bin spacing is not a guarantee of artifact-isolation accuracy.
Display-analyzer FFT	Spectrum sizes follow the performance setting: 2,048 / 4,096 / 8,192. The stereo scope analyzers use 1,024 / 2,048 / 4,096.
Latency	Latency accounting includes spectral buffering, the one-sample de-click stage and limiter look-ahead. Offline output is trimmed back to the requested duration; comparison uses an aligned dry signal.
Monitoring startup	The graph requests a buffered monitoring latency hint of 0.08 seconds. Silent DSP warm-up and gain ramps reduce abrupt starts; the device’s actual output latency is not fixed by that hint.
Scene processing	Smart Repair calibration/validation uses a reusable local Worker running the embedded DSP in block-sized steps. It reports sample-based progress, supports cancellation and checks returned lengths/rates/channel data.
Full Master render	The full offline Master path uses OfflineAudioContext and the same embedded repair-chain processor, followed by selected peak/level processing and WAV encoding.
Final-WAV reader	A dedicated PCM16/PCM24/float32 RIFF reader constructs a playback buffer from the prepared WAV without re-encoding it or routing it through the live Master chain.
Audio sample rates	Master rendering follows the working AudioBuffer rate. Separation resamples for its 44.1 kHz path; other analysis/model paths use their own preprocessing. Final normalization is not a resampling control.
WAV structure	RIFF/WAVE PCM or IEEE float output, validated sizes/alignment, metadata comment and numerical checks. Large allocations remain subject to browser memory and RIFF size limits.

Implementation: Signal-processing engine and spectral implementation · Live/offline spectral profiles, graph and transport · Final loudness engine, presets and controls · Shared encoded-WAV preview/export asset

19 / RUNTIME & DEPLOYMENT

Local-first, with clearly separate optional services

The main application ships as an HTML file containing its interface and restoration code. Some features deliberately load resources or use platform capabilities outside that file.

Capability	What the build needs / checks
Local core work	JavaScript, browser file access/decoding and adequate memory. Direct file opening is possible for portions of the workflow, but not a promise that every advanced API works there.
Live DSP and native offline render	Web Audio / AudioWorklet support in a permitted context. HTTPS or a trusted localhost deployment is the documented route.
Advanced visualizers	WebGL2 availability. The independent defect-map path does not require the live analyzer renderer.
Neural separation	Model and runtime availability, workers, memory, appropriate execution provider and a compatible hosting context.
GPU model paths	WebGPU and required device limits/features. A menu selection does not establish GPU compatibility.
Threading / isolation	The build reports cross-origin isolation and documents COOP/COEP hosting for reliable advanced inference. A plain static server may not provide these headers.
Model storage	IndexedDB and/or OPFS where available. Refused or unavailable storage can require memory-only operation or reloading assets.
PWA / offline installation	A compatible deployment with the companion manifest and service-worker resources. The HTML alone is not a complete installable offline package.
Direct file-system actions	Capability-checked file picker APIs, with standard picker/download fallbacks for applicable actions.
MIDI and gamepad	Relevant browser APIs, a connected device and any required permission/interaction.
Remote URLs	Reachable resources and compatible cross-origin policies. Audio/project launch URLs are fetched rather than implicitly granted local-disk access.
External verification / server transcription	User-selected network actions or a configured local HTTP service, separate from the embedded DSP.

Performance controls

Laptop, Normal and Ultra preview profiles change spectral preview size and analysis/display workload. The machine benchmark can suggest a mode. The Performance / Platform panel reports AudioWorklet load, a recent DSP peak, preview FFT size, latency, UI frame interval and JavaScript heap information when the browser exposes it. Offline HQ/Ultra choices remain separate from the preview mode.

Data boundaries

Core restoration works with loaded audio locally in the browser. Model and package downloads still contact their hosts. MuScriptor Server sends transcription work to the configured service. The optional Suno Credentials actions upload the selected audio only after the application’s explicit consent step. This guide itself contains no network-backed search, tracker, model loader or audio upload.

Distribution

Distribute the application with the companion deployment files required by the features you advertise, retain the source and model notices, and verify the target browser/OS/GPU combinations. The current local-launch bundle is one deployment option; an HTTPS-hosted release is another. Model rights, external service availability and storage budgets are separate from the fact that the main interface is a single HTML file.

Implementation: Live/offline spectral profiles, graph and transport · Long-track planning and source-separation engines · Platform services, projects, autosave, gamepad and corpus · Supported launch switches and startup execution

20 / INSPECTION & REPORTS

Keep evidence separate from artistic judgment

The workstation includes diagnostic and developer-oriented surfaces for inspecting files, documenting a process and testing behavior. They are tools to run, not pre-issued certifications.

JSON diagnosis report

Export source/analysis context, repair profile, current settings, references and related diagnostics for inspection outside the interface.

Scientific HTML report

Build a report with source/render hashes, before/after measurements, curves, validation and DSP state. It describes the rendered derivative and recorded settings, not an independent listening assessment.

Provenance Inspector

Inspect original container bytes, chunks/atoms, codec framing hints, tags, encoder strings and possible Suno/C2PA/chirp markers. Copy the inspector JSON or refresh it.

Generation Lab

Compare the recording’s acoustic signature with factory historical references or a locally learned corpus. Overlay current/top/all families; add an analyzed source, import/export a corpus or clear it. The local corpus is capped at 200 entries.

Synthetic DSP benchmarks

Exercise bypass/null behavior, tonal preservation, noise cleanup, phase-swish, persistence/tonal tracking and de-click tests independently of the loaded song.

Developer QA Lab

Run validation checks for DSP numerical behavior, determinism, render alignment/silence, WAV structure, peak safety and splitter planning. Stress-test current audio and export QA JSON.

Structural planning versus live inference

A long-track structural test checks segment coverage and memory planning for a 6:30 input. A separate live certification action requires a source of at least six minutes and a complete neural run. Its result records engine, output count, shortest stem duration, finiteness and elapsed time. A successful plan does not imply that the model has actually run.

Optional source/derivative credentials comparison

The build provides actions to open the Suno verifier, test original source bytes, render/test a repaired WAV, run a paired source-versus-repaired protocol and clear results. It presents a consent notice and a 100 MB-per-request limit for its configured remote API. Availability and the verifier’s outcome depend on that external service.

SRL does not use a credentials verdict to tune the DSP or optimize removal of provenance signals. The original source C2PA manifest is not copied as though it authenticated newly rendered audio. Presence or absence of a local marker is not a cryptographically verified authorship claim.

Implementation: Measurement, reference and analysis logic · Supported launch switches and startup execution · Platform services, projects, autosave, gamepad and corpus · Long-track planning and source-separation engines

21 / LAUNCH REFERENCE

Open SRL in the state your workflow needs

The launch console and URL options configure an application session. The PowerShell-style visual interface parses SRL options; it is not an operating-system shell.

Switch	Purpose
`--help`	Show this launch-option reference and open the PowerShell-style console.
`--lang <code>`	Set interface language: en fr de es pt it da fi is no sv nl ru zh ja ko ar he la pl cs el tr hu ro uk hi.
`--theme <id>`	Select one of 16 themes: carbon ocean cobalt violet rose amber emerald teal arctic copper crimson royal matrix impulse crt minimal.
`--layout <id>`	Set workspace layout: studio wide compact classic.
`--desk <id>`	Choose full workstation or five-knob desk: full five.
`--level <id>`	Set usage level: novice advanced expert.
`--help-mode <id>`	Set integrated help mode: context inline off.
`--perf <id>`	Set preview performance mode: laptop normal ultra.
`--mode <id>`	Open master, splitter, or stem mode.
`--file <url>`	Load an accessible audio URL or relative resource at startup.
`--project <url>`	Load an accessible .srl project URL at startup.
`--action <id>`	Run after loading: none play analyze smart validate.
`--profile <id>`	Set the repair profile hint.
`--preset <id>`	Load a factory preset.
`--wet <0-100>`	Set the global processed mix percentage.
`--splitter <id>`	Select splitter engine: auto demucs bsroformer.
`--stems <2\|4\|6>`	Select requested splitter output count.
`--commercial-safe <0\|1>`	Toggle commercial-safe defaults.
`--developer <off\|qa\|full>`	Developer surfaces: qa mounts the commercial QA Lab and long-track splitter certification; full also opens technical diagnostic panels for this launch only.
`--qa <0\|1>`	Compatibility shortcut: --qa 1 enables developer QA mode; --qa 0 disables it.
`--tour <0\|1>`	Open the guided tour after startup.
`--console <0\|1>`	Open the launch console.
`--reset-ui <0\|1>`	Ignore saved UI preferences for this launch.
`--args "..."`	Alternative query parameter containing a shell-style list of switches.

Example URL query strings

?lang=en&layout=studio&level=advanced&theme=carbon
?mode=splitter&splitter=demucs&stems=4
?lang=en&tour=1
?developer=qa
?file=audio%2Fmy-track.wav&action=analyze

Append a query string to the actual deployed application URL. Source and project URLs must be accessible from that deployment. A startup play request remains subject to browser playback policies, permissions and the application’s busy/lock state. Preferences can otherwise be saved between visits.

Implementation: Supported launch switches and startup execution

22 / QUICK REFERENCE

Choose the tool for the job

These practical entry points tie the capability catalogue to common tasks without implying that every recording needs every processing stage.

Task	Relevant SRL tools	What to check
A chorus gets bass-heavy and loses articulation	Analyze density drift; Sub Control, De-Mud, signed spectral matrix, self-reference and scene automation.	The quiet sections should not become thin or unnaturally bright.
Metallic shimmer or watery upper-band motion	Shimmer Suppressor; spectral texture/noise/swish controls; defect map and Delta.	Cymbals, sibilants, harmonics and musical transients must not be mistaken for defects.
A narrow source opens up or collapses under density	Spatial-intent diagnosis; bidirectional M/S; low-band phase and mono guards.	Compare both stereo and mono; do not impose width on intentionally narrow material.
You need independent component control	Optional neural separation, StemLab role analysis, mix/duck controls, residual/glue.	Stem sum versus source, leakage, phase and the returned Master input.
You need editable notes in another music tool	Stem transcription, quality/timing/sensitivity controls, individual or combined MIDI export.	Wrong notes, merged articulations, tempo assumptions, velocities and channel-limited pitch bends.
You are preparing a social-media listening file	Final normalization destination, peak ceiling, selected WAV format and exact final audition.	The achieved level may be peak-constrained; platform encoding can alter the delivered playback.
You are producing several related songs	Batch queue, per-row strategy, album target, shared presets and separate final normalization choices.	Preserve intentional track-to-track level differences and check every finished file.
You need a reproducible processing record	Project/preset, snapshots/history, source/render fingerprints and scientific report.	Keep the original media outside the app; a project does not replace the audio files.

Terms used in this guide

Term	Meaning here
DSP	Digital signal processing: the embedded algorithms that alter audio.
Stem	An audio component, imported separately or estimated by neural separation.
Master input	The current recording feeding Master Repair; it may already contain StemLab processing.
Delta / Δ	The correction difference relative to a latency-aligned source.
Wet	The fraction of the processed contribution used in the Master result.
Side/Mid or M/S	A stereo representation used for balance, width and compatibility decisions.
Crest factor	The relationship between peaks and average signal level, used as one dynamics indicator.
FFT / STFT	Spectral analysis / short-time spectral processing over windows of audio.
LUFS / RMS	The two final-level modes: gated K-weighted perceived-loudness estimate, or average-energy level.
dBTP	The label used for an estimated intersample-peak level. SRL’s meters are not certified laboratory meters.
ABX	A blind comparison with known A/B references and an unidentified X.
ONNX / Safetensors	Model asset formats used by the optional neural execution paths.
WASM / WebGPU	Browser execution routes for CPU-oriented compiled code or GPU computation.
OPFS / IndexedDB	Browser-managed local storage used by selected caches and saved-session features.
PWA	The install/offline application packaging supported by a suitable deployment with companion resources.

Scope boundaries

SRL works on existing audio. It is not presented here as a song generator, a full arrangement editor, a guaranteed lost-detail recovery system, an infallible stem separator or note transcriber, an author-identification system, a watermark-removal utility, a certified loudness meter or a universal platform-delivery validator. MIDI previews do not preserve source instrument timbres. Neural execution and long-track performance depend on assets and hardware.

Implementation: DSP catalogue and initial parameter values · Tonal/drum transcription and MIDI serialization · Final loudness engine, presets and controls · Application scope and workflows

23 / EXPOSED CHOICE INDEX

Every main dropdown, in one reference

This appendix lists the runtime dropdown choices in the documented build, including their internal option values. Duplicated tutorial mirrors and repeated macro routing slots are omitted; they refer to the same underlying settings.

Pre-processing separator engine

Auto · HTDemucs for 2/4 stems: auto
HTDemucs · balanced 2/4 stems: demucs
BS-RoFormer-SW · heavier / 6 stems: bsroformer

Pre-processing output count

4 stems · Vocals / Drums / Bass / Other: 4
2 stems · Vocals / Instrumental: 2
6 stems · + Guitar / Piano: 6

Splitter engine

Auto · choose by mode / machine: auto
HTDemucs · balanced 4-stem: demucs
BS-RoFormer-SW · quality 6-stem: bsroformer

Splitter outputs

4 stems · Vocals / Drums / Bass / Other: 4
2 stems · Vocals / Instrumental: 2
6 stems · + Guitar / Piano: 6

StemLab duck key

Off: Off
Vocal: Vocal
Drums: Drums

Transcription audio source

Auto · cleaned, no humanize: auto
Treated StemLab stem: treated
Original separated/imported stem: raw

Transcription engine

SRL Stem Consensus · browser: srl
Hybrid SRL + MuScriptor Web · ONNX/WebGPU: hybrid-web
Hybrid SRL + MuScriptor Server: hybrid
Hybrid SRL + MuScriptor Native · WebGPU: hybrid-native
MuScriptor Native · WebGPU: muscriptor-native
MuScriptor Web · ONNX/WebGPU: muscriptor-web
MuScriptor · local server: muscriptor

MuScriptor Web precision

Auto · FP16 WebGPU / INT8 WASM: auto
FP16 · WebGPU: fp16
INT8 · WASM: int8
FP32 · compatibility: fp32

MuScriptor Web execution

Auto · GPU / worker fallback: auto
WebGPU: webgpu
WASM · proxy Worker: wasm

Transcription quality

Fast · single decode: fast
High · 3-way consensus: high
Max · 5-way consensus: max
Ultra · 11-way consensus + session reconstruction: ultra

MIDI timing

Preserve performance: off
Smart groove · soft 1/16: smart
Quantize 1/16: 16
Quantize 1/8: 8
Quantize 1/4: 4

Tonal sensitivity

Conservative: conservative
Balanced: balanced
Sensitive: sensitive

Drum sensitivity

Conservative: conservative
Balanced: balanced
Sensitive: sensitive

Stem WAV encoding

32-bit float: 32f
24-bit: 24
16-bit: 16

Analyzer focus

All 4 analyzers: all
Vectorscope: phase
Spectrum: spec
Oscilloscope: wave
Defect map: repair

Spectrum content

Source + processed: both
Difference · processed − source: difference
Processed only: processed
Source only: source

Analyzer visual style

Scientific: scientific
Studio: studio
Spectacular: spectacular

Editable defect lane

Bass: bass
Veil: veil
Glare: glare
Space: space
Texture: texture
Shimmer: shimmer
Click: click
Phase: phase
Transient: transient

Master WAV encoding

32-bit float: 32f
24-bit: 24
16-bit: 16

Normalization destination

Balanced listening · −16 LUFS: balanced
YouTube · suggested −14 LUFS: youtube
Facebook · suggested −14 LUFS: facebook
Instagram · suggested −14 LUFS: instagram
TikTok · suggested −14 LUFS: tiktok
TV Europe · EBU R128 · −23 LUFS: tv
Radio · EBU programme reference · −23 LUFS: radio
CD preparation · suggested −16 LUFS: cd
Vinyl preparation · suggested −18 LUFS: vinyl
Custom: custom

Normalization measurement

Perceived loudness (LUFS): lufs
Average level (RMS): rms

Repair profile hint

Profile hint: Auto: auto
Known source: v2 / chirp-v2-xxl-alpha: v2legacy
Known source: v3 / chirp-v3-engine-i: v3
Known source: v3.5 / chirp-v3p5-engine-s-2: v35
Known source: v4 / chirp-v4-engine-d: v4
Known source: v5 Cover / chirp-crow-t1: v5cover
Known source: v5 / chirp-carp-up-c-1: v5carp
Known source: v5.5 / chirp-fenix: fenix55
Known source: v6-mini: goose
Known source: v6-mini + MaxMode: gooseMax
Known source: v6 / chirp-hawk: hawk
Known source: v6-Pro + MaxMode / chirp-hawk: hawkProMax
Generic / unknown generator: generic

Factory preset

Conservative restoration: conservative
v2 / XXL legacy reconstruction: v2legacy
v3 / engine-I density reconstruction: v3
v3.5 / adaptive legacy repair: v35
v4 / narrow-core preservation: v4
v5 Cover / crow conservative repair: v5cover
v5 / carp adaptive repair: v5carp
v5.5 / chirp-fenix density repair: fenix55
Legacy collapse repair: oldCollapse
Chirp-hawk high-band guard: hawkGuard
v6-Pro + MaxMode density texture stabilizer: hawkProMax
v6-mini density repair: gooseMini
v6-mini MaxMode density + width repair: gooseMax
Adaptive full repair v29: adaptive
Spectral artifact repair: spectralSafe
Neutral / all effects flat: neutral

Legacy intensity

Legacy · Authentic: authentic
Legacy · Reconstruction: reconstruction
Legacy · Rescue: rescue

Performance panel profile

Laptop · lighter scopes/analysis: laptop
Normal · balanced: normal
Ultra · denser preview/analysis: ultra

Acoustic-signature overlay

Top matches: top
All families: all
Current only: current

Local corpus family

v2: v2legacy
v3: v3
v3.5: v35
v4: v4
v5 Cover: v5cover
v5: v5carp
v5.5: fenix55
v6-mini: goose
v6-mini MaxMode: gooseMax
v6: hawk
v6-Pro/Max: hawkProMax
Other / custom: generic

MIDI CC target

Wet: wet
Macro · Clarity: clarity
Macro · Dynamics/Punch: punch
Macro · Space: space
Macro · Texture: texture
Macro · Safety: safety
Selected DSP parameter: selectedParam
Switch · A/B: ab
Switch · Delta: delta
Switch · Play/Pause: play

Workspace layout

Studio: studio
Analyzer: wide
Console: compact
Classic Rack: classic

Desk mode

Advanced Studio: full
Simple Repair: five

Visual theme

Carbon Mint: carbon
Deep Ocean: ocean
Cobalt Studio: cobalt
Violet Night: violet
Rose Quartz: rose
Amber Console: amber
Emerald Room: emerald
Teal Precision: teal
Arctic Glass: arctic
Copper Tape: copper
Crimson Stage: crimson
Royal Blue: royal
Matrix Green: matrix
Impulse Tracker / Demoscene: impulse
CRT Analyzer: crt
Dark Minimal: minimal

Usage level

Novice: novice
Advanced: advanced
Expert: expert

Parameter control style

Knobs · universal precision: knobs
Hybrid · knobs + sliders: hybrid
Sliders · accessibility: sliders

Help presentation

Context bubbles: context
Inline explanations: inline
Off: off

Preview performance

Laptop: laptop
Normal: normal
Ultra preview: ultra

Interface density

Comfortable: comfortable
Compact: compact
Dense: dense

Implementation: DSP catalogue and initial parameter values · Supported launch switches and startup execution · Interaction layers, navigation and preferences · Tonal/drum transcription and MIDI serialization · Final loudness engine, presets and controls

24 / DOCUMENTATION & CREDITS

Build-specific, not promise-based

A complete capability description is most useful when its scope, dependencies and evidence remain visible.

This presentation documents ML_SunoRepairLab_v41_9_CORRIGE.html. Descriptions are grounded in its embedded implementation, active interface and integrated help. Active runtime behavior is used where older static labels remain in the file—for example, the profile-dependent preview FFT and the encoded-WAV preview path. The page does not claim to have independently certified every engine or every supported environment.

Application build: SRL v41.9
Documentation language: English
Document edition: 3 October 2026
Source size: 8,701,463 bytes
Source SHA-256: 54deaa2f2d4fe950c8fb2f7e3a2c635b35a6844fd02b9298ed916e0551f7622f

Implementation cross-reference

Source area	Exact anchor / symbol	Location in source
Application scope and workflows	`const TUTORIAL_STEPS=[`	Lines 17,928–17,943
DSP catalogue and initial parameter values	`const effectDefs = {`	Lines 13,867–13,946
Signal-processing engine and spectral implementation	const WORKLET_CODE = String.raw`	Lines 13,967–14,128
Live/offline spectral profiles, graph and transport	`async function ensureAudio(){`	Lines 14,131–14,204
Measurement, reference and analysis logic	`async function estimateTruePeak4x`	Line 15,130
Smart Repair, validation and multi-objective calibration	`async function smartRestore(){`	Lines 15,359–15,448
Final loudness engine, presets and controls	`function createFinalLevelEngine()`	Lines 12,254–12,577
Long-track planning and source-separation engines	`function longTrackPlan(`	Line 15,690
StemLab mixing, import and channel controls	`function roleDefaults(`	Lines 16,190–16,282
Tonal/drum transcription and MIDI serialization	`function stemMidiEngineMode`	Lines 16,282–17,849
Shared encoded-WAV preview/export asset	`/* ---------------- v41.8: one encoded WAV`	Lines 18,938–19,119
Snapshots, presets and parameter assistance	`function updateSnapshotUI()`	Lines 14,426–14,460
Interaction layers, navigation and preferences	`function applyUx()`	Line 14,601
Supported launch switches and startup execution	`const LAUNCH_OPTIONS=[`	Lines 13,738–13,833
Platform services, projects, autosave, gamepad and corpus	`/* ------------------------------ v36/v37 platform`	Lines 18,502–18,554
Playback owner and processing-stage identity	`function srl419Descriptor()`	Line 18,851
Batch strategy and album adjustments	`function batchCentroidHz(`	Lines 15,449–15,475
Credits and third-party notices	`Programmer / original application`	Lines 11,177–11,337

Credits

Programmed by Maxime LORRAI (aka KurtZ/lapiNIC).
© 2026 Maxime LORRAI. Original application code, interface, presets and documentation are reserved to their author except where third-party licences apply.

Cyber artist / beta tester: Sylvain Selmane — Shades.

The application credits Demucs/HTDemucs and its research authors, demucs-web, ONNX Runtime Web, bs-roformer-web and related architecture work, Spotify Basic Pitch, Tonejs/Midi, MuScriptor (Kyutai × Mirelo), C2PA technical references, Suno Credentials and ESM module hosting. Model weights can have different terms from the software that loads them. Consult and retain the relevant notices and obtain any permissions needed for the intended distribution.

Use the application only with content you are entitled to process and distribute. Keep independent backups. The application’s own risk and warranty notices remain applicable; this product guide does not replace them.

Screenshots show the actual English interface using an illustrative synthetic diagnostic signal. No user recording, neural-separation success claim or music before/after benchmark is embedded in this guide.

Repair the sound.Keep the control.

For music creators

For producers and editors

For technically minded listeners

Measure and locate

Repair and protect

Separate and remix

Recover editable notes

Compare and finish

Work and document

Nine editable defect lanes

Self-reference and the Reference Atlas

Sub Control

De-Click / De-Crackle

De-Mud

Defect-Keyed Signed Spectral Matrix

Self-Reference Adaptive Match EQ

Presence Recovery

De-Harsh

Fizz / De-Ess

Spectral Intelligence Engine

Shimmer Suppressor

Adaptive De-Resonance

Safe Air / Exciter

Bidirectional Density M/S

Legacy Phase Coherence Guard

Legacy Harmonic Reconstructor

Transient Restorer

Conditional Line Guard

Look-Ahead Peak Safety

Multi-objective calibration

Representative-scene validation

Chain Order Advisor

A/B and Wet

A/B/C comparison

Delta listening

Blind ABX

Loop and hotspot work

Explicit player switching

Long-track processing and memory

Handoff and cancellation

Analyze stems

Smart Stem Mix

Deterministic humanization

Role-aware spectral duck

Residual / glue

Return to Master

Choose which audio is transcribed

Tonal analysis and four quality settings

Timing, expression and drums

Exports and audition

Optional model paths in detail

WAV encoding and correction outputs

Normalization: on by default, TikTok selected

How final level is calculated

Batch and album tools

Snapshots, undo and preset exchange

.srl projects and autosave

Repair Assistant macros

Rebase macros

Five-knob repair desk

Expert macro matrix

All 16 factory presets

16 visual themes

27 interface languages

Web MIDI control

Gamepad

Performance controls

Data boundaries

Distribution

JSON diagnosis report

Scientific HTML report

Provenance Inspector

Generation Lab

Synthetic DSP benchmarks

Developer QA Lab

Structural planning versus live inference

Optional source/derivative credentials comparison

Example URL query strings

Terms used in this guide

Repair the sound.
Keep the control.