NG Processor User Manual

This manual explains how to install, configure and operate NG Processor. It is written for the people who run a station, whether or not they have a technical background.

The focus is practical operation: audio input and output, the Input Balance Level, presets, Easy Tune, the meters and health indicator, file processing and the NG Encoder plugin. The signal chain is described to the depth an operator needs to make good decisions; the full technical parameter catalogue is documented inside the application, per parameter.

The following procedure is sufficient for a normal live installation between playout and encoder.

The main window combines preset selection, transport and a complete meter bridge without a separate monitoring screen.

2.1Preset selector and transport

2.2INPUT and OUTPUT meters

Horizontal peak meters from -30 to 0 dBFS with red peak-hold markers. INPUT shows the feed after the Input Level Correction and before the internal reference line-up, so the meter describes the programme the processing stages actually receive; OUTPUT shows the processed audio at the configured ceiling. In BYPASS both meters follow the same corrected samples.

The NG PROCESSOR title line carries the recognised programme class (MUSIC, SPEECH or the blend) on the right, and the health bar at the bottom carries a CPU figure for the processing engine next to the audio format. The CPU figure is the share of real time the engine needs: the time spent processing a block divided by the duration of that block, on the audio thread, independent of how many cores the machine has; a Predictive engine with full projections sits well below 100 % on a modern machine, and a value that creeps towards 100 % is the first warning of an underpowered or overloaded computer.

2.3AGC RIDER

Two bipolar rows, MAIN and BASS, show the net gain the AGC currently applies: gain grows to the left in yellow (up to +24 dB), attenuation to the right in white (to -4 dB); zero is the bright mark. On a soft record or a fade-out the bar walks left; on a hot master it sits a little right of zero. When the bar stops moving left, the rider has reached the standing-gain ceiling of the preset.

In Predictive a third row, SHADOW GAIN, sits under BASS on the same scale. Its dark-grey bar is the net MAIN gain of the independent ordinary AGC that runs alongside the Predictive rider (5.7): same settings, same input, but never the Predictive programme gain, fade hand-back or a Shadow takeover. It is a reference, not a correction and not the output of a second Classic chain; comparing it with MAIN shows how far the plan is away from what an ordinary rider would do. While the reservoir primes or no data is available, the row stays dimmed without a value; BYPASS clears it and STOP lets it fall back gently. In Classic the row is absent and the shared scale sits directly under BASS; the independent reference rider only exists in Predictive.

2.45-BAND DYNAMICS and CAP

Five vertical gain-reduction meters for LOW, L-MID, MID, H-MID and HIGH, each with a peak-hold marker. Healthy operation is a few decibels of movement; a band that sits permanently deep indicates a source or voicing problem, not a reason to drive harder. Below the bands, six activity bars: RESTORE, BALANCE, DYN, AIR, GUARD and BASS. RESTORE shows the combined restoration activity of Peak Restore and Revive (8.8).

Beside each band column stands a narrow bipolar meter for the Band Leveler (8.7): the slow correction that band currently receives, ±6 dB around a centre line. Green grows upward for a boost, red downward for a cut. When the active preset does not use the Band Leveler, or during BYPASS, the narrow meters stay dimmed and empty.

A separated sixth, blue column labelled CAP shows Content Aware Processing (chapter 8.6): how much of the preset's processing is currently being withheld because the record itself is already over-processed. It fills from the top like the band meters; an empty column means the record receives the full preset. When the active preset has Content Aware Processing disabled, the column stays in place but is greyed out.

The meter header carries a programme badge - MUSIC, MIXED or SPEECH - taken from the exact Music/Speech blend the DSP is using (below 25%, 25-75%, above 75%). MIXED is a real transition region, not a hidden switch.

2.5FINAL CONTROL

LIMIT L and LIMIT R show the final look-ahead limiter's gain reduction on a 0-8 dB scale with peak hold. Dense programme shows one or two decibels of steady movement with deeper excursions on transients only.

2.6Health bar

The right side of the bar shows the audio format in use, for example AUDIO 48.0 kHz · STEREO.

2.7Engine switch and the Future Timeline

The CLASSIC / PREDICTIVE pair next to TUNE selects the engine for the next start; it is locked while LIVE. In Predictive the input/output strip is replaced by the Future Timeline (5.4), and the status line reports the reservoir fill, NOW • MONITORING, ATTENTION or FAIL-SAFE. The meters and the health bar behave as in Classic; note that the output meters follow the delayed output, five seconds behind the input.

Open File > Settings… while processing is stopped.

Calibrate measures the feed for you. Stop the processing, choose the input and sample rate, play representative music at the normal source level and press Calibrate. The input is opened without starting any output, 20 seconds of real frames are measured (at least 10 seconds of usable programme are needed) and the robust median programme level is turned into a recommended Input Balance Level on the whole-dB grid, taking the chosen Input Level Correction and the selected preset's internal line-up into account. The button reads Cancel while measuring.

Only Save applies the result; Cancel on the window keeps the previous value. Silence, a clipped feed, a device problem or an implausible result are refused with an explanation. Calibrate does not touch hardware volume, L/R balance, AGC drive, ceilings or presets, and it is not a perceptual loudness alignment: play the material your station actually broadcasts.

Every preset is a JSON file in the Presets folder. Nothing is compiled into the executable except Bypass. Editing a preset never requires a new build, and a station can version and back up its sound with ordinary file tools.

4.1The shipped presets

Eight factory presets ship: four for the Classic engine and four for Predictive. Personal presets are never renamed or overwritten. Further personalities are delivered as .json files that can be copied into the Presets folder of any installed version and appear after Preset > Reload presets from disk.

All factory presets use an Output Trim of 0 dB. This is not loudness normalisation: the different voicings and densities still produce different average levels.

Every factory preset also carries a Revive character (8.8): Balance none, Airwave, Pure, Velvet and Vertigo Gentle, Coast and Horizon Medium, Impact Strong. All six relative Easy Tune controls (Warmth, Bass, Punch, Density, Brightness, Width) start at zero on every factory preset: the factory sound is the baseline, and a slider shows how far you have moved away from it. AGC HEADROOM is an absolute value and is not reset.

4.2Factory presets and your own

A factory preset shows a star and the word FACTORY. The moment you change anything in it, it becomes your own preset: the star disappears, the file stays where it is, and the sound is exactly what you made of it. You can always go back by reloading the shipped file (12.3). Give your own presets their own names so a factory file is never overwritten.

4.3Preset menu

C:\Program Files\NG Processor\
NGProcessor.exe
NGProcessor.config
Presets\
  Balance.json        ★ FACTORY
  Airwave.json        ★ FACTORY
  Pure.json           ★ FACTORY
  Coast.json          ★ FACTORY
  Impact.json         ★ FACTORY
  Velvet.json         ★ FACTORY
  Horizon.json        ★ FACTORY
  Vertigo.json        ★ FACTORY
  Station Sound.json  custom

The factory presets are listed first when present, all other files alphabetically, Bypass last. Values in preset files are stored with two decimals.

4.4Predictive presets

Switching Predictive presets on air: the new preset is prepared in the background, its name appears on the Future Timeline from the moment it is being prepared, and the hand-over happens in one activation window that includes the 350–500 ms crossfade. The audio no longer waits for a presentation deadline; the five-second buffer is the look-ahead, not a switching delay. Rebuilding a Predictive engine from the kept history still takes a few seconds; the timeline shows when it is ready.

Predictive uses its own preset family, stored as target profiles in the same Presets folder. Four factory profiles ship: Impact, Velvet, Horizon and Vertigo (see 4.1). Predictive presets are separate from the Classic presets (Balance, Airwave, Pure, Coast and Bypass); a Classic preset cannot be selected while the Predictive engine is active, and vice versa.

Your own Predictive presets are saved from Easy Tune through the Preset menu; the status line confirms with Predictive preset saved: name. A custom Predictive preset is complete only after it has been saved once by NG Processor; the NG Encoder plugin refuses incomplete profiles.

4.5Generate from WAV

Preset > Generate from WAV… makes a preset from a recording. Point it at a processed / reference WAV, the sound you want to learn, and optionally at the original / unprocessed recording of the same programme. The engine that is selected (Classic or Predictive) is the engine the preset is generated for. The function is available on registered copies only; Predictive generation also needs the Predictive entitlement.

• Reference only gives an estimate: the reference is measured (spectral balance, loudness, dynamics, stereo) and a preset is built from a neutral factory base towards those measurements. It is a starting point, not a match.
• Reference plus original gives a fit: the two recordings are matched on content, the original is rendered through the real selected engine with candidate presets, and the output of the complete chain is scored against the reference on density, crest, short-term consistency, level transfer, nine spectral regions, stereo width and fade behaviour. A bounded search (at most six rounds, eighteen renders) keeps only changes that measurably improve the score, and a separate control passage that was not used for searching decides whether the result stands; if it does not, the first candidate is kept. The result is labelled fitted (all primary measures within tolerance on fit and control), approximate or estimate. Approximate is not the same as equal: another processor's short-term behaviour is not always reachable with settings.
• What you get. A new custom preset next to your other presets, plus a .analysis.txt with the measurements. Nothing existing is overwritten and the WAVs are not copied; only the learned reference curve is stored. The results window shows the changed settings with their JSON paths, and with a pair an Output comparison tab with reference, initial output, chosen output, tolerance and deviation, separately for fit and control. Load preset opens the result through the normal preset selection; Easy Tune then starts at zero around the generated sound.
• Material. WAV, mono or stereo, PCM 16/24/32 or float. At least a few minutes of real programme with music and transitions; longer is better. The fit needs the same programme on both sides. For your own station that is the playout feed recorded alongside the old processor, the typical case when replacing hardware. For another station you will not have its feed; the nearest substitute is the records it played, from your own library, in the same order: the matcher pairs on content, not on time, and skips jingles, talk and anything that does not match. Expect fitted for your own station or an NG preset, and approximate for another make of processor: tonal balance, density and stereo come close, but another processor's short-term behaviour is not always reachable with settings, and the report says so. For Predictive the processor warns before long jobs (five minutes or 100 MiB and more), because the fit renders the original several times with the full look-ahead engine.

NG Processor has two engines. Classic is the real-time chain described in the rest of this manual. Predictive is the same native chain, driven by a supervisor that hears exactly five seconds of programme before it reaches the processing. It classifies what it hears as source events — a fade, a hard cut, a quiet intro, silence — on evidence and admission criteria, and plans gain, tone, dynamics and transient behaviour ahead of time, sample-accurately. Event help is applied only when the evidence meets the preset's criteria; an ambiguous passage is left to the ordinary AGC. The price is five seconds of latency.

5.1What Predictive does

Audio enters a reservoir of exactly 5.000 seconds. Every 250 ms a background analysis reads the whole reservoir, recognises programme events (fades, hard transitions, intros, breakdowns, reverb tails, silence, speech) and writes a plan onto the timeline: six control lanes — programme gain, dynamics authority, spectral tilt, transient authority, finalizer pressure and stereo width — that the native chain applies at the exact sample they were planned for. The last 250 ms before “now” are committed and not rewritten, so a decision is not revised at the last moment.

In practice this is what you hear: a long outro is lifted as a whole because the supervisor already knows the next track starts in 3.2 seconds; that first beat is caught on the sample rather than smeared into the fading track; a quiet intro gets a bounded assist; and silence is left alone. The ordinary AGC keeps levelling within a track; Predictive only takes over the exceptional moments it can see coming. How far it may lift is bounded by the preset and by the AGC HEADROOM budget (6.2).

A second, independent Shadow AGC runs alongside the plan on the same source: an ordinary AGC rider that never receives the Predictive lanes, so it always knows what the level would have been without them. It audits the plan against that reference and corrects on a separate lane when the two disagree (5.7).

5.2Latency, and when to use it

Predictive adds exactly five seconds to the Classic chain latency. That makes it ideal for streams, automation feeds, file processing and the NG Encoder plugin, where nobody is waiting for the output. It is not suitable for a presenter monitoring the processed signal. For live studio monitoring, use Classic.

5.3Switching and priming

The CLASSIC / PREDICTIVE switch sits next to TUNE in the toolbar. The engine can only be changed while stopped; while LIVE the status line reads Stop LIVE before switching between Classic and Predictive engines. After START in Predictive the Future Timeline reads BUFFERING 41% (the fill percentage counts up) while the reservoir fills; the output is silent for those five seconds, then processing begins and the timeline reads NOW • MONITORING. The last engine you used is remembered for the next start.

5.4Reading the Future Timeline

In Predictive mode the input/output strip of the main window becomes the Future Timeline: the programme waveform from one second in the past (PAST 1s) through NOW into the seconds ahead (+1s, +2s, +3s of the five-second reservoir), with the planned actions drawn on it. NOW • GAIN +6.0 dB is the programme gain applied at this instant. The waveform is drawn after the Input Level Correction, with 0 dBFS at the full height of the strip; audio pushed above full scale by a positive correction is only clipped visually against the edge, while the absolute value stays readable on the INPUT meter. Already played history keeps the correction that applied at that moment, also while BYPASS is pressed.

• GAIN LIFT / GAIN TRIM — the programme-gain lane is raising or lowering the level ahead (a fade lift, a hard-transition step, a silence guard).
• TRANSITION — a hard boundary between two programme items; gain is stepped here, never smoothed across it.
• FADE OUT / FADE IN — a recognised fade, drawn as a sloping band over the analysed stretch; the caption sits at the side where the fade starts. A fade out is what the GAIN LIFT lane acts on.
• BRIGHTEN / DARKEN — the spectral-tilt lane is opening or closing the top end.
• DYN DRIVE / DYN OPEN — the dynamics lane is asking the multiband stage for more or less authority.
• ATTACKS — the transient lane is giving attacks more room.
• SILENCE / SPEECH — what the supervisor believes it is looking at. While the programme is silent the SILENCE marker stands on NOW; a silence that begins later in the window is marked where it starts.
• SHADOW GAIN +x.x / −x.x dB — in orange, top centre, only while a Shadow correction is actually being applied. A negative value is the Shadow AGC taking over from a lift that proved wrong (at most two seconds, 5.7); a positive value is the Future Shadow safety net lifting a soft new start that the plan did not claim (up to +6 dB, 5.7). Its recent course is drawn as a thin trace in the PAST area. No caption means the plan and the ordinary AGC agree; the SHADOW GAIN row under the AGC meters shows the independent rider at all times.

5.5Health, ATTENTION and FAIL-SAFE

NOW • MONITORING is normal operation. PREDICTIVE • ATTENTION means the processing carries on but something needs a look; the health line says what, for example that the input device delivered audio faster than real time after a restart. PREDICTIVE • FAIL-SAFE means the look-ahead analysis stopped: NG Processor then quietly falls back to the Classic chain, the audio never mutes, and the Predictive help returns by itself as soon as the analysis recovers. Both situations are written to the log. ATTENTION clears by itself after a few stable seconds; if it keeps coming back, see chapter 13.

5.6Availability

Predictive is part of the license. Without it the CLASSIC / PREDICTIVE switch is not shown and NG Processor runs the Classic chain, which is complete in itself. The NG Encoder plugin has its own arrangement (chapter 11).

5.7The Shadow AGC

The Predictive plan is decided five seconds ahead. Between the decision and the moment it plays, the programme can do something the plan did not count on: the source comes back louder after a fade, a hand-over turns out to be a breakdown, a presenter starts talking over the end of an outro. The Shadow AGC is the engine's second opinion for exactly those moments. It is an ordinary AGC rider with the same settings as the live one, fed the same audio, that never receives the Predictive lanes. It therefore always knows what the ordinary rider would have done on its own; the SHADOW GAIN meter (2.3) shows that value.

The Shadow AGC does not ride along continuously. A negative Shadow takeover creates no extra positive gain; it returns control to the ordinary AGC. It only acts when earlier Predictive or fade support turns out to be wrong, and only after proof: the source recovers above what the lift assumed, confident speech arrives after a music fade, or the independent rider settles on a stable, clearly different level while the source no longer falls. An ordinary AGC lag or a Quiet Intro on its own is never a reason to intervene.

The Future Shadow. The same safety net also works the other way round: when a soft new song starts and the plan has not recognised it as an intro, the Shadow lifts it a little (at most 6 dB) until the normal automatic gain has caught up. On the timeline that reads SHADOW GAIN +x dB. It never adds to help that is already running.

The end of a lifted fade is handed back to the ordinary AGC explicitly. When the source itself dictates a fast programme-gain release at that moment (a hard next-track onset, typically 50-60 ms), the authorised hand-back takes that release over as well, which is designed to prevent an audible drop before the ordinary rider carries the level; the gain and limiter meters show whether that succeeded on a given transition. When a lift ends, only the AGC gain that the lift had displaced against the shared ceiling is given back, and only as far as the source now needs it; the level does not fall away and then recover.

TUNE opens Easy Tune for a factory preset or a preset derived from one: broad musical tone controls plus the three dynamics decisions an operator is most likely to need. Zero is exactly the preset, changes preview live, and a double-click returns a slider to zero.

Every preset selection or Easy Tune move is applied live: the engine hands over to the new settings through a 350 ms crossfade, also when the change alters the processing latency, so nothing is deferred until STOP / START. If a replacement engine fails to prepare, the previous one simply stays on air. BYPASS continuity is preserved across the hand-over.

Easy Tune moves are smoothed per sample over 70 ms on the audio clock, so a slider sweep glides instead of stepping. Tone controls move in 0.1 dB steps, dynamics macros in 1% steps; the mouse wheel steps one unit, Shift-wheel ten. The tone macros are real broad tonal shaping - clearly audible - without changing compressor detector calibration.

On the factory presets the six relative controls start at zero. Zero is the factory sound; a value shows your offset from it, and RESET returns to the factory baseline.

6.1Saving an Easy Tune result

Saving from Easy Tune stores the factory baseline together with the eight slider values. The star and FACTORY badge disappear (the file is no longer untouched), but the preset keeps its Easy Tune workflow: reopening shows the sliders exactly where they were saved and never applies the macro twice. Any change that cannot be explained by the sliders (for example a hand-edited preset file) ends Easy Tune ownership; such a preset can still be selected and used, but TUNE no longer offers Easy Tune for it.

6.2Predictive Easy Tune

With the Predictive engine selected, TUNE opens Predictive Easy Tune: character and AGC headroom for the selected profile. Changes are heard live; double-click a slider to reset it, RESET returns all of them, CANCEL restores the original profile and APPLY keeps the result. The six character controls run from −100% to +100% in 5% steps, where 0% is exactly the profile; AGC HEADROOM runs from 0 to 16 dB in 0.5 dB steps.

About AGC HEADROOM. It is one budget for all positive gain: it sets the maximum gain of the ordinary AGC and the ceiling of the Predictive programme-gain lift. As a simplified example: while Predictive is lifting a fade by 8 dB on a 12 dB budget, the AGC has 4 dB left; gain cuts do not create extra room. In practice requests can displace one another within the shared budget, so judge the net MAIN gain and the programme-gain lane separately rather than adding figures. Impact ships at 12 dB, Velvet and Horizon at 12.5 dB and Vertigo at 14 dB. The slider is a ceiling, not a drive: how much of it is used on a given fade depends on how deep the fade is and how sure the supervisor is.

6.3Expert

Easy Tune covers the character of a preset with a handful of controls. Expert, the button next to RESET in both Easy Tune windows, opens every parameter of the selected preset. It is the same editor for Classic and Predictive; in Predictive it also carries the planner sections.

• Sections. On the left a tree: General, Input (level, filters, phase corrector), Restoration (Peak Restore, Revive, Source Repair, Dynamics Restore), Tone & Stereo (Dynamic EQ, Tone / EQ, Band Leveler, Stereo, De-esser, Subharmonics), AGC & Level (AGC, Bass, Content-aware, Speech / Music), Multiband (Global and the five bands) and Output (the finalizer). Predictive adds Character / Headroom, Programme targets, Fade / Outro, Fade cutoff, Quiet intro and Safety limits.
• Grid. Every setting shows its exact JSON name as it appears in the preset file, its plain-English name, and the value with its unit, range and step. True/false and choice settings use a drop-down. A value outside its range, or one that conflicts with a coupled setting, is refused as a whole, with the reason.
• Help. Selecting a setting shows a plain-language explanation of what it does, what the typical range means and how it interacts with its neighbours, plus the exact path for people who edit the JSON by hand.
• Search. The search box finds settings across all sections by name, JSON name or keyword, in English and Dutch.
• Apply, Cancel, Reset. Apply hands the edited settings to the live engine without a gap, like a preset change. Cancel restores the state the editor opened with; Reset returns to the values the preset was loaded with. After an Expert edit the preset counts as yours: the factory star disappears until you reload the factory file.

Expert is for engineers who know what they are reaching for. For the daily sound, Easy Tune is the intended control surface, and the factory presets are tuned so that Expert is rarely needed.

The AGC is a two-sided rider around the internal target level. Audio above the target is pulled down (up to the Max Reduction); audio below it is pulled up (up to the standing Max Gain). Small movement inside the floating window passes untouched; outside it a windowed ratio corrects part of the error - 3.6:1 in Airwave corrects about 72% - slowly enough that the five-band section owns the musical movement.

7.1States

Short breaks. A drum break, a stop-time or a dramatic pause is not a quiet passage. The AGC recognises a source stop (a fall of more than 12 dB within 100 ms against the two-second median, followed by near-silence or further decay) and holds its gain for up to 2.5 seconds instead of recovering towards the idle gain. When the music returns at a comparable level the rider simply continues; a markedly hotter return still gets the normal fast entry. The gate also closes faster, on a 50 ms peak-hold detector, while opening keeps the slower, qualified detector. The aim is no dip after the break and no pumping on breaks, in Classic and Predictive alike.

7.2Outros, intros and the next track

On a fade-out the rider walks toward its standing-gain ceiling; Airwave allows +16 dB. After a genuine gap a bounded quiet-intro reserve is armed for a soft intro and withdrawn quickly when the loud part arrives. When the next track starts on top of a lifted outro - without passing through the gate - the rider does not restart. A level jump of at least the Sudden Jump threshold first arms an immediate transient freeze (1050 ms in Airwave), so stings, drum fills and jingles are simply not chased. If the jump persists for 60 ms and is large, a bounded Fast Attack Assist - common to Master and Bass, 180 ms attack, 650 ms release - takes up to 2.5 dB of the entry while the slow rider follows at its normal attack; the assist withdraws as soon as the rider has caught up. The 900 ms Acquisition state is reserved for start-up and for a genuinely hot start after a gated gap, so a mistaken transition decision is bounded instead of audible.

A separate gradual-fade guard (120/800 ms programme detector) recognises a sustained fade: once the slow envelope has dropped 4 dB over at least 700 ms and this holds for 200 ms, further Master/Bass recovery is limited to 3 dB until the fade ends. Long outros are still lifted, but the rider no longer stores excessive gain that would slam the next track; stable quiet intros and abrupt quieter programme are unaffected.

7.3Reading the AGC meter

The bar sits around zero on nominal audio, left of zero when quiet material is being lifted, right of zero on hot material. A bar pinned at the left edge for a whole track means the record is quieter than the preset's standing gain can compensate: a preset decision (Max Gain, Idle Gain), not a reference-level problem.

8.1Restoration (before the AGC)

Peak Restore reconstructs clipped peaks in hot masters, optionally with a spectral deep-restore pass (roughly 210-230 ms of extra latency when enabled in real time, depending on sample rate), guarded by a confidence law so natural flat crests are left alone. Source Repair notches mains hum with harmonics, the 15.625 kHz line whistle and a 19 kHz pilot only when detected. All three are transparent when nothing is detected. Clipping is detected from an amplitude histogram rather than from bit-exact flat plateaus, so clipped material that has been resampled, dithered or level-shifted on its way in — as it usually has when it arrives through a virtual audio cable or a stream — is still recognised.

8.2Five-band dynamics

Crossovers at 100, 400, 1600 and 6400 Hz. Each band combines a progressive compressor with a fast per-band limiter and band-limited clip-error control. The Drive sets how hard the section works as a whole; the Output Mix per band is the tonal colour knob (post-compressor, not regulated back); Common-Mode Calm prevents the five bands from pumping together on the beat. Healthy operation is a few decibels of gain reduction, not a permanent deep hold.

8.3Bass processor

An 18 dB/oct LF contour adds weight; a clean pre-limiter and a deep peak guard share a single reduction budget, so that the loudest bass note of a chorus is not pushed softer than its neighbours. The Low band's compressor levels sustained heavy bass before it reaches the fast guards.

8.4Tone, air and spectrum balance

The three-band parametric EQ sits in front of the multiband; about half of a static EQ change survives on the long-term spectrum because the band regulates back - use the band output mix for station voicing. The air lift is a second-order shelf with its own guard; the Balance calibration places it at 13.5 kHz. Spectrum Balance 2 slowly normalises persistently dull or bright sources with deadbands and sudden-change protection.

8.5Natural Dynamics, stereo and speech

Dynamics Restore G2 returns crest and punch to flattened masters within a bounded budget. The stereo enhancer widens within headroom limits with mono bass below 170 Hz. A de-esser and a low-octave synthesiser are available and off by default.

The Music/Speech detector combines centre dominance, syllabic modulation and voice-band balance, with a rhythmic-periodicity veto so mono or centre-heavy music is not mistaken for speech. Its decision is a continuous 0-100% blend (about 2.5 s attack, 4.2 s release) with hysteresis: SPEECH is entered at 56% after a one-second dry-speech qualification, left at 40%, and every state is held for at least five seconds. Downstream stages - restoration, LF contour, multiband and finalizer - interpolate along that blend; nothing switches hard. The gentler speech personality is only applied when Alternate Speech Personality is enabled in the preset (off in every factory preset); Content Aware Processing fades out on speech regardless. Mode can be pinned to Music or Speech for pure formats - a non-stop music station is best served by Music, which also removes any chance of a false SPEECH verdict on centre-heavy records.

8.6Content Aware Processing

Content Aware Processing works in Classic and in Predictive alike. Content Aware Processing answers one question about the record that is playing: how over-processed is it already? The more flattened, clipped or brickwalled the master, the less NG Processor adds on top. It only ever withholds processing; it never adds any, and it never changes level directly.

The analysis runs on the untouched source, before Peak Restore, phase rotation and every other stage, so NG Processor's own work cannot influence it. Over a three-second window it measures the peak-to-loudness ratio of the source (a flat master sits at 8-9 dB and below, open programme at 12 dB and above), strengthened by flat tops - runs of samples parked at the ceiling - and by the time the waveform spends within 1.4 dB of its peak. Those two only amplify a verdict the ratio has already started, so a playout limiter clipping the odd peak on a dynamic record does not trigger anything.

A record is judged by its densest loud section: the window scores feed a memory that rises in about 4 s and releases in about 8 s, and stronger reduction is only learned from windows within 4 dB of the recent loud level, so verses and breakdowns do not add reduction. A programme gap resets the decision; on speech the reduction fades to zero.

The blue CAP column in the 5-BAND DYNAMICS chart (2.4) shows the withheld fraction. On an open, well-mastered record it stays empty; on a flattened dance master it fills within a few seconds of the first drop and stays filled through the breakdown. Setting all three reductions to 0, or disabling the section, leaves every record with the full preset.

8.7Band Leveler

The wideband AGC sets the level of a record; it cannot change its colour. A record that is dull, bass-heavy or harsh stays that way, and the fast five-band section only rides the short-term movement within it. The Band Leveler adds a slow gain regulator for each of the five bands, placed after the AGC and before the fast band dynamics. It measures how each band relates to the whole over a few seconds and draws that relationship calmly towards the preset's target colour, so records sound as though they belong to the same station without being turned into each other.

• Slow and bounded. Corrections move at about a decibel per second, with a deadband around the target and a maximum boost and cut per band. Neighbouring bands can only differ by a limited amount, so the leveler never draws extreme curves. On average it changes colour, not level.
• Target colour. Each preset defines where the leveler aims, per band. A neutral target (as in Pure) keeps the preset's own voicing; Horizon, for example, aims for more low-mid and mid body and a calmer upper mid. Because the leveler regulates towards its target, a static EQ change placed before it is partly regulated back: the target is the place where the tonal balance of a leveler preset is set.
• Gates and speech. The leveler holds while the programme is gated or a band is almost empty, digital silence stays silence, and its strength is reduced on speech so voices are not re-equalised.
• Quiet Lift. Optionally the leveler lifts short quiet stretches - speech pauses, soft intros - by a bounded amount, judged against the recent programme level rather than a fixed target.
• Fades and Shadow. During a Predictive fade lift the colour regulation keeps tracking normally, while additional Quiet Lift is withheld so a lifted outro is not lifted twice. During a real Shadow takeover (5.7) the whole band vector holds until the takeover is complete.
• Predictive. In the Predictive engine the leveler also uses the tonal outlook of the next five seconds, so a colour change at a hand-over is prepared rather than chased.

The factory presets Pure, Coast, Velvet, Horizon and Vertigo use the Band Leveler; Balance, Airwave and Impact do not. Its activity is shown by the narrow meters beside the band columns (2.4).

8.8Revive

Peak Restore (8.1) rebuilds peaks that were clipped off. Many modern masters are also limited and compressed so hard that the movement inside the music is gone: the difference between a snare hit and the space around it has been squeezed out. Revive gives that movement back. It works directly after Peak Restore and in front of the AGC, in several frequency bands, and reads the untouched source to decide how much dynamic structure each band should regain. The rest of the chain then processes a record that breathes again instead of a flat brick.

• Three characters. Strong, Medium and Gentle set how much dynamic structure is restored; Off switches the stage out. The character is part of the preset. Seven of the eight factory presets have Revive on (Balance is the neutral no-Revive reference); material that is already dynamic needs little help and is left largely alone.
• Latency. An active Revive adds 896 samples (20.3 ms at 44.1 kHz, 18.7 ms at 48 kHz). LIVE keeps one stable presentation delay that already includes this budget, so presets with and without Revive can be switched on air without a jump or click. Revive runs at 44.1 and 48 kHz; at other rates the stage is transparent.
• What you see. Revive has no meter of its own: its activity is included in the RESTORE activity bar (2.4). A RESTORE bar that moves on loudness-war masters and stays quiet on open recordings is normal operation.
• Levels. Revive can raise short peaks above the incoming level; the AGC, the five-band dynamics and the final limiter downstream handle that as usual, so the output ceilings are unaffected.

The final stage is a clean look-ahead ladder: a look-ahead limiter with a sliding-minimum gain envelope, a 4× oversampled clipper with overshoot compensation for the last fraction of a decibel, an optional crest shaper for broadcast density, and adaptive distortion guards that back the chain off when the audio would start to sound processed.

Choose File > Open WAV to process, select a WAV file and choose where to save the processed result. The current working preset is applied with the offline path, which renders exactly the same audio as the live path and compensates the processing latency, so the output stays aligned with the input.

• Use file processing to audition a preset change on a known track before applying it on air.
• Compare a processed file against the original at matched loudness; the processed file will be louder by design.
• Sample rate follows the file; the presets behave identically at every supported rate.
• The progress bar shows the measured processing speed (audio seconds per second of processing). CANCEL beside the progress stops the job: an existing output file stays untouched and temporary files are removed. If the result had already been written, a late Cancel does not undo it.
• File processing writes only the processed WAV; no diagnostic or preset side files are created next to it.
• With the Predictive engine and a file of five minutes or 100 MiB and more, a Yes/No warning (No by default) appears first: the look-ahead engine takes longer than real time on such files. Classic and short files start directly.

NGProcessorPlugin.dll makes the same processing engines available inside NG Encoder. It is one self-contained native DLL: place it next to NGEncoder.exe; the host loads it from its own folder. Do not copy NGProcessorCore.dll or any managed NG Processor assembly beside the encoder — the plugin needs neither; it does need a Presets folder beside the encoder (11.1).

11.1Presets in the host

The plugin has no presets built in except Bypass. It loads every preset from the Presets folder next to NGEncoder.exe: copy the factory files there (Balance, Airwave, Pure and Coast for Classic; Impact, Velvet, Horizon and Vertigo for Predictive, listed as Predictive · Impact and so on) together with your own .json presets. A custom Predictive profile must be complete — save it once in NG Processor — or the plugin refuses it. The default is Bypass; if the selected preset file is missing or unreadable, the plugin runs Bypass, reports it and loads no substitute.

11.2Latency

The plugin reports its latency to the host: the Classic chain latency, plus exactly five seconds when a Predictive preset is selected. NG Encoder aligns metadata and metering to it; selecting a Predictive preset in a live stream therefore adds five seconds of delay to the air signal.

11.3Licensing

The plugin runs on an NG Encoder with a license that includes the plugin option, a demo license as well as a regular one: tick Enable NG Processor and it runs. Without a license the plugin is not available. Updating it later is a matter of replacing one file while the encoder is stopped.

12.1Setup program

The setup program installs per machine to Program Files\NG Processor, creates a Start menu entry and an optional desktop icon, and grants users write access to the installation folder because configuration and presets live there. It checks the registry for Microsoft .NET Framework 4.8 and installs the bundled redistributable silently only when it is missing (Windows 10 version 1903 and later and Windows 11 already include it); a required reboot is reported at the end of setup.

12.2Updates

Reinstalling or installing a later build works in place. A running NG Processor is closed and restarted, NGProcessor.exe is replaced, and the shipped preset files are installed only where no file of that name exists yet - an edited factory preset stays as you left it. Configuration, custom presets and the log are never touched; the uninstaller leaves the presets in place.

12.3Restoring a factory preset

To return an edited factory preset to its shipped state, delete or rename its file in the Presets folder and run the setup program again, or copy the file from another installation. The certificate inside the file re-validates automatically and the star returns.

12.4Portable use

NGProcessor.exe, NGProcessorCore.dll (from the same release) plus the Presets folder in any writable folder is a complete installation. Keep one folder per instance when running more than one processor on a machine.

12.5Registration and license

NG Processor runs without a license, with restrictions: one copy at a time, every five minutes the audio drops sharply for ten seconds, and processing a WAV file stops after 60 seconds. An unregistered copy reminds you of this at start-up; Continue closes the reminder, Enter registration code… opens the License window.

Open Help > License, enter the registration code you received and press Activate. The computer needs an internet connection at that moment; afterwards NG Processor keeps its license up to date by itself, and a computer that is offline for a while simply keeps working. Should the license ever lapse, the program tells you so in the License window; Refresh now checks it again, for example after a change in your customer account.

The Predictive engine is part of the license (5.6). The NG Encoder plugin is available with a demo or regular NG Encoder license that includes the plugin option (chapter 11).

13.1Only Bypass is listed

• The Presets folder next to NGProcessor.exe is missing or empty. Copy the preset files in and use Preset > Reload presets from disk.

13.2A factory preset lost its star

• The preset file was edited and no longer matches its certificate. This is intended; the preset is now yours. See 12.3 to restore the shipped file.

13.3No meter movement

• Confirm the WASAPI input in Settings, the upstream routing and the Windows level and mute state. Audio I/O is locked while processing; stop first to change devices.

13.4PROCESSOR HEALTH: ATTENTION

• Read the detail: INPUT BUFFER · x% FULL indicates live buffer pressure from the audio device or the machine. Check CPU load, device drivers and other real-time audio software. The state clears itself when the pressure is gone.

13.5The AGC does not lift a quiet record or a fade

• Classic: watch the AGC meter: pinned at its left limit means the preset's standing gain is exhausted. Use Airwave, or raise Max Gain / Idle Gain in the preset. Do not lower the Input Balance Level for this purpose (chapter 3).
• Predictive: work through the chain in order. Detection: does the Future Timeline show FADE OUT or QUIET INTRO at all? Admission: a candidate without a lift was recognised but not admitted (chapter 5). Request: GAIN LIFT shows how much was asked. Limit: the shared AGC HEADROOM (6.2) caps the ordinary AGC and the lift together. Output: a lift that is applied can still be taken back by the limiter; check the FINAL CONTROL meters before raising any gain.

13.6The first beat after a lifted outro hits the limiter

• Expected briefly: the look-ahead limiter catches the very first transient, the Fast Attack Assist takes up to 2.5 dB within about 200 ms and the slow rider follows at its normal attack. If the level stays hot for seconds, the entry exceeded what the assist may take: this is by design (a bounded response is preferred over a re-acquisition dip); lower Master Attack slightly or check that the preset's Sudden Jump Threshold is not set unreasonably high, which would also disable the transient freeze.

13.7Presets sound the same

• Confirm the version in the title bar matches the build you expect, and that the files in Presets are the ones you think they are (a custom file with a factory name shows without a star).

13.8Processing a WAV fails

• The file must be PCM or IEEE-float WAV. Check NGProcessor.error.log next to the executable for the reason.

13.9The CAP column fills on most records

• The library is loud, modern product with clipped or limited masters: Content Aware Processing is doing its job. If the station sound becomes too polite, lower the three Reduction values in the preset rather than disabling the section; if the column also fills on open, classic masters, check that no limiter or clipper in the playout chain is flattening the feed before it reaches NG Processor.

13.10The programme badge shows SPEECH on music

• Centre-heavy, sparsely arranged or mono records can qualify as speech for the five-second minimum state. On a non-stop music station set Speech → Mode to Music; the badge then always reads MUSIC and Content Aware Processing never fades out. Stations with presentation should leave Auto and, if the alternate personality is unwanted, keep Alternate Speech Personality off.

13.11The PREDICTIVE button is unavailable

• The installed license token does not carry the predictive entitlement, or the copy is unregistered. Classic keeps working; see chapter 5.6.

13.12Five seconds of silence after START in Predictive

• Normal: the Future Timeline shows BUFFERING nn%. Output starts when the reservoir is full.

13.13“Stop LIVE before switching between Classic and Predictive engines”

• The engine is chosen before START. Press STOP, switch, START again.

13.14PREDICTIVE • ATTENTION with “reservoir overflow: source discontinuity”

• The capture source delivered more audio than time passed — a driver restart, a sample-rate change or a device switch. The reservoir drops the excess and continues; if it repeats, check the capture device and its sample rate.

13.15PREDICTIVE • FAIL-SAFE

• Runtime health became critical, usually because the analysis stopped. All Predictive lanes are neutral and the Classic chain runs on; audio never stops. It recovers by itself; if it stays, note the reason in the log and restart the engine.

13.16The timeline lifts less than the AGC HEADROOM slider allows

• The slider is a ceiling. The planned lift follows the depth of the fade and the supervisor's confidence; a shallow fade gets a small lift however much room you offer.

13.17SHADOW GAIN shows a negative value while GAIN LIFT is active

• That is the Shadow AGC taking over (5.7): the source has proved that the lift no longer fits, and the level is handed to the ordinary AGC. The negative value lasts at most two seconds; after that the ordinary AGC owns the level and the retired fade no longer lifts.
• Compare MAIN with the SHADOW GAIN row (2.3). If the plan regularly lifts where the independent rider sees no reason to, lower AGC HEADROOM or OUTRO lift strength in Predictive Easy Tune.
• A positive SHADOW GAIN on a soft new start is the Future Shadow safety net (5.7): the plan did not recognise the start as a Quiet Intro, so the independent rider lifts it temporarily. If this happens on starts that should have been recognised, check the intro settings (6.2) rather than the Shadow.

• Set the Input Balance Level to the nominal level of the feed and verify with a line-up tone.
• Listen to the chosen preset for a representative broadcast hour on the players your audience uses.
• Confirm the true-peak ceiling against the distribution chain (FM, DAB, streaming codecs).
• Play a soft intro after a gap, a long fade-out into a hard track start, and a jingle on level programme; watch the AGC meter and the final limiter.
• Check that the health bar stays GOOD under normal machine load; note any ATTENTION causes.
• Save the station sound as a custom preset and back up the Presets folder and NGProcessor.config.
• If NG Encoder is used with the plugin, verify that both point at the same preset files.
• Confirm operators know that BYPASS is latency-matched and that Easy Tune zero equals the preset.