# CALM EverClean: Technical Architecture, Signal Flow, and User Guide

## 1. Abstract

CALM EverClean is a real-time audio processor focused on voice, feedback control, dynamic cleaning, and channel-strip processing for live and studio use. The system combines a higher-selectivity spectral/neural path with a zero-added-latency causal path.

The design separates the problem into two processing families:

- **Spectral/neural control with controlled latency**, used by CALM in Studio 21 ms mode.
- **Causal sample-by-sample processing**, used by EverClean, DNS Zero Latency, and CALM Zero Latency.

This architecture makes it possible to work in two very different scenarios: latency-sensitive live sound, where every sample matters, and contexts where 21 ms is acceptable in exchange for more precise analysis.

This document explains the public architecture of the plugin, its modules, signal flow, latency modes, usage recommendations, and installation instructions. It does not include protected implementation details or sensitive engine information.

## 2. Design Goals

CALM EverClean is designed around six primary goals:

1. **Predictable latency**: Studio 21 ms and true Zero Latency operation.
2. **Simple main controls**: two clear cores, CALM and EverClean.
3. **Dynamic vocal cleanup**: reduce noise, bleed, room tail, and tonal build-up without turning the plugin into a fixed EQ.
4. **Vocal protection**: avoid destroying body, consonants, whispers, or useful air while cleaning.
5. **Live usability**: fast interface, clear metering, stable CPU behavior, and reliable recall.
6. **Robust distribution**: embedded encrypted internal resources, with no external model files for the user to copy.

## 3. General Architecture

The plugin is organized into five functional blocks:

```text
Input
  |
  |-- Mode Router
  |     |-- Studio 21 ms path
  |     |     |-- CALM FFT/neural control
  |     |     |-- Resonance / spectral modules
  |     |     |-- DNS frame-based processing
  |     |
  |     |-- Zero Latency path
  |           |-- EverClean causal cleaning
  |           |-- CALM causal feedback guard
  |           |-- DNS causal split
  |
  |-- Channel Strip
  |     |-- Expanse
  |     |-- Clean EQ
  |     |-- Body
  |     |-- De-esser
  |     |-- Compression
  |     |-- Finisher
  |
  |-- Auto Gain / Output Trim
  |
Output
```

The key design decision is that the plugin does not try to solve every problem with a single processing model. Studio mode and Zero Latency mode have different responsibilities.

## 4. Latency Modes

### 4.1 Studio 21 ms

Studio is the maximum-selectivity path. It uses spectral analysis and neural control to make finer decisions about tonal content, feedback risk, and cleanup.

Characteristics:

- Latency approximately equivalent to a 1024-sample window at 48 kHz.
- Latency display in milliseconds and samples.
- Higher spectral resolution.
- Better context for tonal control, resonances, and feedback.
- Recommended for mixing, broadcast, streaming, theatre, installations, studio work, and live situations where the latency is acceptable.

In this mode, CALM has more temporal and frequency information available. For that reason, Studio is the recommended mode for maximum anti-feedback precision.

### 4.2 Zero Latency

Zero Latency avoids lookahead and does not report latency to the host. The path works causally, using filters and states that depend only on the present and the past.

Characteristics:

- Zero added latency.
- EverClean is active as the main cleaning engine.
- CALM Zero Latency acts as a causal tonal guard.
- DNS Zero Latency acts as causal noise reduction.
- Recommended for monitors, sensitive live use, spoken voice, lavaliers/headsets, and any situation where latency is critical.

The fundamental difference is that Zero Latency cannot "see the future." Its behavior therefore relies on recurrent states, envelopes, causal filters, and tonal-stability detection.

## 5. CALM Engine

CALM is the anti-feedback and tonal-stability core.

### 5.1 CALM in Studio 21 ms

In Studio, CALM uses frame-based spectral information. The signal is analyzed through bands and bins to estimate:

- narrow stable tones,
- temporal stability,
- feedback risk,
- tonal build-up,
- the relationship between vocal energy and resonant energy,
- the amount of dynamic reduction required.

The neural engine should not be understood as an "effect" that draws a fixed curve. It acts as a control estimator. The DSP then performs smoothed reduction to avoid sudden jumps, unnecessary modulation, and damage to the voice.

Relevant controls:

- **CALM On/Off**: enables the anti-feedback core.
- **Strength**: main CALM amount.
- **Turbo**: stronger response for difficult situations.
- **Tonal Lock**: tracks stable tones.
- **Low Tonal Lock**: extends stable-tone tracking into lower frequencies.

Recommended use:

- Start with CALM enabled and moderate Strength.
- Enable Turbo only when the system is close to the limit.
- Use Studio when maximum precision is needed.
- Use Tonal Lock if persistent tones appear even while program material is active.

### 5.2 CALM in Zero Latency

In Zero Latency, CALM switches to a causal narrow-band guard. Instead of using an FFT window with lookahead, it analyzes the signal sample by sample through a narrow-band filter bank.

Conceptual process:

1. Split the signal into narrow bands.
2. Measure envelope and energy in each band.
3. Estimate a dynamic floor per band.
4. Detect tonal dominance and stability.
5. Apply causal cancellation/reduction in suspicious bands.
6. Smooth attack and release to avoid artifacts.

This mode is not intended to replace the full precision of CALM Studio, but it provides a zero-latency safety layer.

## 6. EverClean Engine

EverClean is the zero-added-latency vocal cleaning engine.

### 6.1 Function

EverClean is designed to act on problems that often accumulate in a live vocal channel:

- background noise,
- stage bleed,
- room tail,
- early reflections,
- tonal dirt,
- small build-ups that make the voice feel distant or less intelligible.

The engine works in real time without lookahead. It uses persistent internal states and a causal path to adapt cleanup to the signal.

### 6.2 Mono and Stereo Processing

EverClean can process mono and stereo material.

- In mono, it processes the vocal channel directly.
- In stereo, it preserves the relationship between channels and avoids unnecessarily damaging the image.

For live sound, the most common use case is mono individual vocal channels. For stereo vocal buses or stems, the stereo path maintains output coherence.

### 6.3 Main Control

The main control is **EverClean Strength**.

Suggested ranges:

- **0-25%**: very light cleaning, useful for delicate sources.
- **25-50%**: natural cleaning for spoken voice or singing with moderate noise.
- **50-70%**: strong cleaning, useful for stages with bleed or difficult rooms.
- **70-100%**: aggressive mode. Use carefully and listen to consonants, air, and body.

EverClean should be adjusted by listening in context. If it is pushed too high on an already clean source, it can start to sound over-processed.

### 6.4 Relationship with DNS

EverClean and DNS do not do exactly the same job:

- EverClean targets overall dynamic vocal cleanup.
- DNS targets background noise reduction.

In Zero Latency, both can operate together. Recommended workflow:

1. Adjust EverClean until the voice moves closer.
2. Enable DNS if stationary noise remains.
3. Keep DNS conservative if EverClean is already working hard.

## 7. DNS

DNS is a dynamic noise-reduction engine.

### 7.1 DNS in Studio

In Studio, DNS operates with band-based information and noise-floor estimation. It maintains a history of the minimum/probable noise level and reduces it when the energy falls below a dynamic threshold.

Conceptual process:

1. Analyze energy per band.
2. Estimate noise floor.
3. Compare against a dynamic threshold.
4. Generate gain per band.
5. Smooth attack/release.
6. Protect vocal bands.

### 7.2 DNS in Zero Latency

In Zero Latency, DNS uses a causal five-zone frequency split. Each zone maintains:

- envelope,
- noise floor,
- minimum gain,
- attack/release,
- vocal protection.

Controls:

- **DNS Bias**: controls how strongly the detector leans into reduction.
- **DNS Attenuation**: controls the available maximum reduction.

Recommendation:

- In Basic mode, use Auto DNS.
- In Advanced mode, start with low values and increase gradually.

## 8. Auto Gain

Auto Gain compensates output level after processing. This matters because a cleaning processor can appear to sound "better" simply because it is quieter. Auto Gain helps make comparisons more honest.

Recommendation:

- Keep Auto Gain enabled while testing.
- Disable it only when measuring the true amount of level reduction or when manual gain control is preferred.

## 9. Advanced Channel Strip

Advanced mode turns CALM EverClean into a complete vocal channel strip.

### 9.1 Expanse

Expanse works as dynamic background reduction. It is useful for controlling noise or room sound between phrases without hard chopping.

Controls:

- Threshold
- Depth
- Speed

Recommended use:

- Spoken voice: moderate depth, medium speed.
- Singing: lower depth to preserve breaths and musical tails.

### 9.2 Resonance

Resonance is designed to control narrow resonances. It is especially useful in Studio mode, where more spectral information is available.

Controls:

- Sensitivity
- Depth

Recommended use:

- Raise Sensitivity until resonances are detected.
- Raise Depth only as much as needed.
- Avoid extreme values if the voice begins to lose body.

### 9.3 Clean EQ

Clean EQ provides two areas:

- Focus: weight, punch, warmth, or low-zone control.
- Clarity: presence, air, sheen, or silk.

It is not a replacement for EverClean. It is a tonal finishing section.

### 9.4 Body

Body controls proximity and vocal weight. It can help restore weight if cleaning makes the voice too thin.

### 9.5 De-esser

The de-esser reduces sibilance. It includes wideband, split-band, and spectral modes.

Recommended use:

- Use it after adjusting EverClean/DNS.
- If EverClean is pushed high, revisit de-essing because perceived sibilance may change.

### 9.6 SSL-style Compression

Channel/bus-style compression for stabilizing dynamics.

Recommended use:

- Moderate ratio.
- Auto makeup for a faster workflow.
- Avoid heavy compression before solving severe feedback.

### 9.7 Vocal Finisher

Final color module with drive, opto, and air. Use it at the end, not as a replacement for cleanup.

## 10. Recommended Workflows

### 10.1 Live Vocal with Feedback Risk

1. Select **Studio 21 ms** if latency is acceptable.
2. Enable **CALM**.
3. Raise **Strength** until the system stabilizes.
4. Enable **Turbo** only if necessary.
5. Enable **EverClean** if there is also noise, bleed, or room tail.
6. Enable **Auto Gain** for fair comparison.
7. Adjust the channel strip at the end.

### 10.2 Monitors or Latency-Critical Situations

1. Select **Zero Latency**.
2. Enable **EverClean**.
3. Set EverClean Strength between 25% and 60%.
4. Enable **CALM** if there is tonal/feedback risk.
5. Enable **DNS** if stationary noise remains.
6. Avoid extreme settings if the performer monitors directly through the processed signal.

### 10.3 Corporate Lav/Headset

1. Start in **Zero Latency**.
2. EverClean between 35% and 65%.
3. Enable Auto DNS if there is air conditioning, room noise, or steady background noise.
4. Use gentle Expanse if there is noise between phrases.
5. Use de-essing if the microphone is bright.

### 10.4 Theatre

1. Use **Zero Latency** if monitoring is latency-sensitive.
2. Use **Studio 21 ms** if the system allows latency and more control is needed.
3. Keep EverClean moderate to avoid damaging breaths or performance detail.
4. Keep DNS conservative.
5. Use Body to restore proximity if cleaning makes the voice too thin.

### 10.5 Broadcast / Streaming

1. Use **Studio 21 ms** if latency is not a problem.
2. Use moderate CALM.
3. Use moderate EverClean.
4. Add DNS according to the background noise.
5. Use Clean EQ, de-esser, and compression for the final tone.

## 11. Suggested Starting Points

### Natural Live Vocal

```text
Mode: Studio 21 ms
CALM: On
CALM Strength: 35-55%
Turbo: Off
EverClean: On
EverClean Strength: 30-50%
DNS: Auto / conservative
Auto Gain: On
```

### Zero-latency Speech

```text
Mode: Zero Latency
CALM: On
CALM Strength: 25-45%
EverClean: On
EverClean Strength: 35-60%
DNS: On if needed
Auto Gain: On
```

### Difficult Noisy Vocal

```text
Mode: Studio 21 ms if possible
CALM: On
CALM Strength: 45-65%
Turbo: On only if feedback is close
EverClean: On
EverClean Strength: 50-70%
DNS: On, moderate
Expanse: optional, light to medium
Auto Gain: On
```

### Emergency Feedback-prone Channel

```text
Mode: Studio 21 ms preferred
CALM: On
CALM Strength: 60-80%
Turbo: On
Tonal Lock: On
Low Tonal Lock: On only if low feedback is present
EverClean: use only as needed
```

## 12. Installation

### 12.1 macOS

Included formats:

- AU: `CALM EverClean.component`
- VST3: `CALM EverClean.vst3`
- Standalone: `CALM EverClean.app`

Common install locations:

```text
~/Library/Audio/Plug-Ins/Components/
~/Library/Audio/Plug-Ins/VST3/
/Applications/
```

If macOS blocks the beta because it is not yet notarized:

```bash
xattr -cr ~/Library/Audio/Plug-Ins/Components/CALM\ EverClean.component
xattr -cr ~/Library/Audio/Plug-Ins/VST3/CALM\ EverClean.vst3
killall -9 AudioComponentRegistrar
```

Then restart Logic, LiveProfessor, Reaper, or the relevant host.

### 12.2 Windows

Included formats:

- VST3: `CALM EverClean.vst3`
- Standalone: `CALM EverClean.exe`

Common VST3 install location:

```text
C:\Program Files\Common Files\VST3\
```

If Windows displays a security warning, proceed only if the file comes from the official EverWorks package.

## 13. Recommended Beta Tester Reports

For useful feedback, report:

```text
OS:
Host:
Plugin format:
Buffer size:
Sample rate:
Source:
Mic:
Mode: Studio 21 ms / Zero Latency
CALM Strength:
EverClean Strength:
DNS:
Auto Gain:
What improved:
What got worse:
CPU / glitches:
Would you trust it live:
```

Example of useful feedback:

```text
macOS 14, Logic, AU, 48 kHz, 128 buffer.
SM58 vocal through wedge.
Zero Latency, CALM 40%, EverClean 55%, DNS On.
Reduced room noise and small ringing around 3 kHz.
Above 70% EverClean consonants started to feel processed.
No CPU issue.
Would use live around 45-55%.
```

## 14. Troubleshooting

### The plugin does not appear in Logic

1. Close Logic.
2. Run `xattr -cr` on the AU.
3. Run `killall -9 AudioComponentRegistrar`.
4. Open Logic again.
5. Check Plugin Manager.

### The host only shows mono or the wrong stereo layout

1. Restart the host.
2. Verify that the plugin has been rescanned.
3. Test on a mono track and on a stereo bus.

### It sounds too processed

1. Lower EverClean.
2. Lower DNS Attenuation.
3. Disable Turbo if there is no real feedback risk.
4. Enable Auto Gain to compare levels.
5. Restore body with Body or Clean EQ if needed.

### I do not hear enough reduction

1. Verify that the module is enabled.
2. Raise Strength gradually.
3. Try Studio 21 ms if maximum selectivity is required.
4. Test with a source that actually contains noise, room tail, or feedback risk.

### CPU is high

1. Use Zero Latency if the system needs minimum processing load.
2. Disable advanced modules that are not needed.
3. Increase buffer size for mixing or broadcast.
4. Avoid duplicating heavy chains across many channels unless necessary.

## 15. Distribution Security

The current version embeds EverClean resources in encrypted form inside the binary. On load, the plugin validates the data before initializing the engine. This improves:

- portability,
- integrity,
- resistance to incomplete installations,
- engine version control,
- reduction of external files.

macOS signing and notarization will be added in a later distribution phase.

## 16. Known Limitations

CALM EverClean should not be understood as a miracle tool. Practical limitations include:

- It does not fix poor gain structure.
- It does not replace proper system alignment.
- It cannot remove extreme physical feedback if the system is already out of control.
- It does not turn a heavily contaminated recording into a perfect studio vocal.
- Aggressive settings can affect consonants, air, or body if pushed too far.

Correct use is as an additional control layer, not as the only solution in the system.

## 17. Conclusion

CALM EverClean combines anti-feedback control, zero-latency vocal cleaning, noise reduction, Auto Gain, and an advanced channel strip in a free tool focused on real-world use.

The current architecture is designed for two worlds:

- **Studio 21 ms**, when maximum precision is needed.
- **Zero Latency**, when latency is the priority.

The direction of the project is clear: keep the plugin free, simplify the workflow, and improve cleaning through real feedback from real users.