Glossary

This glossary defines terminology commonly used in capture, profiling, neural audio modeling, reamping, and model playback.

Terminology is not completely standardized across manufacturers and open-source projects. Different platforms may use words such as capture, profile, model, and clone for related but technically different processes.

Where a term has a platform-specific meaning, that distinction is noted.

A

A1

The original family of standard Neural Amp Modeler neural-network architectures, retrospectively named A1 following the introduction of A2.

A1 includes architectures previously known as Standard, Lite, Feather, and Nano.

A1 models remain supported by current NAM software.

A2

The second-generation standard architecture for Neural Amp Modeler, officially released in June 2026.

A2 was developed to improve the relationship between modeling accuracy, computational requirements, training efficiency, and hardware implementation.

See NAM A2 Architecture.

A/B comparison

Direct comparison between two signal paths.

During capture work, this commonly means comparing:

Physical device versus trained model
Original recording versus model output
Two different models of the same equipment

Accurate A/B comparison requires careful level matching.

AIDA-X

An open-source neural amplifier modeling and playback project developed by Aida DSP.

AIDA-X uses its own model architecture and should not be confused with NAM merely because both are open-source neural modeling technologies.

See Open Capture Technologies.

Amp capture

A model representing an amplifier without its physical speaker/cabinet/microphone response.

An amp capture is commonly combined with a cabinet IR or cabinet model during playback.

Also called a direct amp capture in some contexts.

Amp+Cab capture

A capture that includes the amplifier together with a speaker/cabinet signal path.

When created using a microphone, the microphone and its placement are also normally part of the resulting sound.

An additional cabinet IR is generally unnecessary during conventional playback.

See Capture Types.

Algorithmic modeling

Modeling in which a developer designs an algorithm intended to reproduce the behavior of a device or circuit.

This contrasts conceptually with capture technology, where the behavior of physical equipment is measured and a model is derived from those measurements.

Modern products may combine algorithmic and capture-based techniques.

Aliasing

Unwanted frequencies produced when nonlinear processing generates frequency components above the Nyquist frequency and they fold back into the audible range.

Oversampling is commonly used in nonlinear digital audio processing to reduce aliasing.

Architecture

The structure of a neural network.

An architecture defines how neural-network layers and processing elements are arranged.

Examples include:

LSTM
WaveNet-style architectures
NAM A1
NAM A2

Two model files using different architectures are not automatically interchangeable.

Audio interface

Hardware that converts between analog audio signals and digital audio used by a computer.

In a capture workflow, an audio interface may provide:

Training-signal output
Capture return input
Level control
Analog-to-digital conversion
Digital-to-analog conversion

B

Batch training

Training multiple models as part of one automated workflow.

Batch training can be useful when several recordings have been made from:

Different amplifier settings
Different pedals
Different channels
Different pieces of equipment

Bias

An operating condition within an electronic circuit.

In capture discussions, the term is often encountered with tube amplifiers or fuzz circuits whose behavior can change significantly according to bias conditions.

Bias can also have unrelated meanings in machine learning. Context matters.

Block

A processing element placed within a digital signal chain.

Examples include:

Amp block
Capture block
Cab block
IR block
Delay block

Several hardware capture platforms represent captured models as blocks within a larger preset.

Boost

A device that increases signal level and may also alter frequency response or nonlinear behavior.

Boost pedals are common capture targets.

Buffer

An electronic circuit designed to isolate signal stages and provide appropriate impedance relationships.

A buffer can alter the behavior of impedance-sensitive equipment such as certain vintage fuzz circuits.

Its presence or absence can therefore matter during capture.

C

Cab

Abbreviation for speaker cabinet.

In modeling systems, Cab can also refer to a digital cabinet model.

See Cabinets and IRs.

Calibration

The process of establishing a known relationship between digital signal level and analog signal level.

In capture systems, calibration can help ensure that a model receives approximately the same operating level during playback that the physical equipment received during measurement.

See Gain Staging and Calibration.

Capture

The measurement process used to characterize physical audio equipment for creation of a digital model.

The word may also be used informally for the resulting model.

Platform terminology varies.

For example:

Neural DSP — Neural Capture
Line 6 — Clone
Kemper — PROFILE
TONEX — Tone Model
NAM — model

Capture data

Recorded measurement information used to train or construct a model.

Capture data is not necessarily the finished model.

Conceptually:

capture data → training/modeling process → model

Capture signal

A specially prepared audio signal sent through physical equipment during the measurement process.

Also called:

Training signal
Stimulus
Excitation signal
Test signal

Different capture systems use different signals.

Capture signal chain

The complete physical path through which the capture signal travels.

For example:

interface → reamp box → amplifier → load box → interface

or:

capture hardware → amplifier → cabinet → microphone → capture hardware

Everything within the measured signal path can potentially influence the resulting capture.

See Capture Signal Chain.

Capture type

The portion of a signal chain represented by a model.

Common capture types include:

Pedal
Amp
Amp+Cab
Preamp
Complete rig

See Capture Types.

Clone

Line 6's term for a model created using Line 6 Proxy.

Current Proxy Clone categories include:

Amp+Cab
Amp
Preamp
Distortion

The word clone may also be used generically in discussion, but Clone has a specific meaning within the Proxy ecosystem.

Cloud training

Model training performed on remote computing infrastructure rather than entirely on the user's local computer or capture hardware.

Examples include workflows used by:

Neural Capture V2
Line 6 Proxy

Cloud training can provide greater computational resources but requires access to the associated online service.

Colab

Common abbreviation for Google Colaboratory.

Colab provides cloud-hosted Python notebook execution and has been widely used for neural-model training.

NAM and several open-source audio-modeling projects have provided Colab-based training workflows.

Conditioned model

A neural model that receives one or more control parameters in addition to the audio signal.

Conceptually:

audio + control value → model → output

This can allow one model to represent behavior across multiple physical control positions.

Contrast with snapshot model.

Convolution

A mathematical operation widely used in digital signal processing.

Cabinet impulse-response playback typically uses convolution to apply the measured response contained in an IR to an audio signal.

Cortex Cloud

Neural DSP's online service for Cortex content.

It is used for sharing and managing Cortex material and performs the cloud-based training required for Neural Capture V2.

Cortex Control

Neural DSP's desktop application for managing Cortex hardware.

It also participates in the Neural Capture V2 creation workflow.

CPU

Central Processing Unit.

In capture playback, CPU requirements influence:

Number of models that can run simultaneously
Buffer size
Latency
Oversampling
Available effects processing

Different neural architectures can have substantially different CPU requirements.

CustomTone

Line 6's community service for sharing compatible presets and Proxy Clones.

D

dB

Decibel.

A logarithmic unit used to describe ratios such as signal level or gain.

The meaning depends on the reference being used.

dBFS

Decibels relative to Full Scale.

A digital level measurement.

0 dBFS represents the maximum digital level available before numerical clipping in a conventional fixed reference system.

Values below full scale are expressed as negative numbers.

dBu

A voltage-based audio level measurement referenced to 0.775 volts RMS.

Calibration information in some capture systems may use dBu to describe the relationship between digital and analog levels.

Device under test

The physical equipment being measured.

Often abbreviated DUT.

Examples include:

Amplifier
Pedal
Preamp
Compressor
Complete rig

Direct capture

A capture made without a physical speaker/microphone response.

For an amplifier this commonly means taking the output through an appropriate load box or direct-output system.

A cabinet model or IR is normally required when playing a direct amp capture through a full-range monitoring system.

Direct Injection box

See DI box.

DI box

Direct Injection box.

A device commonly used to interface an instrument-level, high-impedance source with a balanced microphone-level input.

A DI box and reamp box perform broadly opposite interface functions, although specific equipment may provide additional capabilities.

DSP

Digital Signal Processing or, depending on context, the processing resources available for digital signal processing.

Hardware modelers often describe limits in terms of available DSP.

DUT

See Device under test.

E

Embedded system

A dedicated computing platform designed to perform a specific function.

Neural-model players can run on embedded processors inside:

Pedals
Amplifiers
Multi-effects processors
Dedicated model players

Efficient neural architectures are particularly important for embedded use.

Excitation signal

See Capture signal.

F

Full-range system

A playback system intended to reproduce a broad frequency range relatively neutrally.

Capture users often use full-range monitors or PA speakers when playing models that already include cabinet simulation.

FRFR

Full Range, Flat Response.

A term commonly used for speaker systems intended to reproduce modeled guitar rigs without adding the strong tonal filtering of a traditional guitar cabinet.

In practice, no loudspeaker is perfectly flat.

Fuzz

A strongly nonlinear distortion effect.

Some fuzz circuits are particularly challenging capture targets because their behavior can depend on:

Source impedance
Input level
Guitar volume
Bias
Playing dynamics

G

Gain staging

Management of signal level throughout an audio chain.

In capture work, gain staging is important both for avoiding clipping/noise and for reproducing the operating conditions under which nonlinear equipment was measured.

See Gain Staging and Calibration.

Ground lift

A facility designed to interrupt an audio ground connection to help eliminate ground-loop hum.

A proper audio ground lift should not be confused with defeating protective mains safety grounding.

Never remove protective mains grounding as a method of eliminating hum.

Ground loop

An unwanted current path created by multiple grounding connections.

Ground loops commonly produce hum or other interference in capture and reamping systems.

H

Hardware capture

A capture workflow in which dedicated hardware performs some or all of the measurement process.

Examples include:

Kemper PROFILER
Quad Cortex
Helix Stadium

The model training itself may occur locally or in the cloud depending on the platform.

Hidden layer

An internal processing layer within a neural network.

The size and organization of hidden layers are part of the model architecture.

I

Impedance

Opposition to alternating current, including audio signals.

Impedance relationships can affect the behavior of:

Guitar pickups
Pedals
Reamp devices
Amplifier inputs
Audio interfaces

Some equipment, particularly certain fuzz circuits, can respond strongly to source impedance.

Impulse response

See IR.

Inference

Using an already-trained neural network to process new input.

In capture terminology:

training creates the model.

inference runs the model.

Real-time capture playback is neural-network inference.

Input level

The signal level presented to a device or model.

With nonlinear equipment, input level affects behavior as well as output volume.

Changing input level can alter:

Distortion
Compression
Sustain
Dynamics
Cleanup

IR

Impulse Response.

A representation of the response of a linear system.

In guitar modeling, IRs are commonly used to represent:

speaker + cabinet + microphone

An IR is fundamentally different from a nonlinear neural amplifier model.

See Cabinets and IRs.

J

JSON

JavaScript Object Notation.

A human-readable structured data format.

Some neural model formats use JSON or JSON-based structures.

The fact that two model formats use JSON does not make them compatible.

The NAM Model File Format is JSON-based.

L

Latency

Delay between an input signal entering a system and the corresponding output being produced.

Latency matters in two different capture contexts:

Recording/training

The recorded output must be correctly aligned with the training input.

Playback

Excessive processing latency can make real-time playing uncomfortable.

See Troubleshooting Captures.

Level matching

Adjusting two signals so they have closely comparable perceived level.

Level matching is essential when comparing:

Model versus physical device
Two models
Two capture platforms

A slightly louder signal can easily be perceived as better.

Linear system

A system whose behavior can be represented using linear signal-processing methods.

Speaker/cabinet/microphone responses are often approximated as linear systems and represented using IRs.

Distorting amplifiers and pedals are nonlinear and generally require more complex modeling.

Load box

A device designed to provide the electrical load required by an amplifier while allowing its output to be recorded or processed without relying solely on a physical loudspeaker.

Load boxes are commonly used when creating direct amplifier captures.

Some load boxes also provide:

Speaker simulation
IR processing
Attenuation
Line outputs

These functions should be configured appropriately for the intended capture type.

LSTM

Long Short-Term Memory.

A recurrent neural-network architecture designed to process sequential information.

LSTM architectures have been used in several neural audio-modeling projects, including GuitarML systems.

M

Machine learning

A broad class of computational methods in which a system learns relationships from data rather than relying solely on explicitly programmed rules.

Several modern capture technologies use machine learning to reproduce nonlinear audio equipment.

Metadata

Information stored with or associated with a model.

Useful capture metadata may include:

Creator
Equipment manufacturer
Equipment model
Capture type
Equipment settings
Input calibration
Output calibration
Training information
Capture date

See NAM Model File Format for NAM-specific metadata.

Mic capture

A capture whose measured path includes a microphone.

For an amplifier rig this usually means:

amplifier → speaker → microphone

The microphone and its position become part of the resulting model or captured response.

Model

A digital representation used to reproduce the behavior of another system.

Within capture technology, a model is the result of training or profiling measurements of physical equipment.

The word is also used more broadly for conventional algorithmic amp and effects models.

Context is important.

Model architecture

See Architecture.

Model conversion

Transforming a model intended for one architecture or playback engine into another representation.

Conversion is not necessarily lossless.

A converted model should not automatically be described as executing the original architecture natively.

See NAM Playback.

Model file

A file containing the information required by a compatible playback engine to run a trained model.

A model file is not necessarily the same thing as the original capture recording.

Model player

Software or hardware capable of executing a trained model.

Examples range from desktop plugins to dedicated pedals and multi-effects processors.

Modeling

The broad process of creating a computational representation of physical equipment.

Modeling can include:

Circuit modeling
Behavioral modeling
Machine-learning modeling
Profiling
Capture technology

N

NAM

Neural Amp Modeler.

An open-source deep-learning system for modeling nonlinear audio equipment.

NAM also commonly refers to the wider ecosystem surrounding the technology.

See Neural Amp Modeler.

.nam

The conventional file extension for a Neural Amp Modeler model.

A .nam file uses the NAM Model File Format.

Not every neural amplifier model is a NAM model.

NAM A1

See A1.

NAM A2

See A2.

Native playback

Playback in which software or hardware executes the model architecture directly rather than first converting it to another proprietary model representation.

See NAM Playback.

Neural Capture

Neural DSP's proprietary capture technology used in the Cortex ecosystem.

Current generations include:

Neural Capture V1
Neural Capture V2

See Neural Capture.

Neural network

A computational structure consisting of interconnected processing elements whose parameters are learned from data.

Neural networks can approximate nonlinear relationships and are therefore useful for modeling audio equipment such as amplifiers and pedals.

Nonlinear system

A system whose output is not simply a scaled or filtered version of its input.

Examples include:

Distorting amplifier
Overdrive
Fuzz
Compressor
Saturating preamp

Neural capture technology is particularly useful for reproducing nonlinear behavior.

Normalization

Adjustment of signal or data level according to a defined reference.

Normalization can mean different things in recording, training, and model playback contexts.

It should not automatically be assumed to be equivalent to calibration.

Nyquist frequency

Half the digital audio sample rate.

For example, at a 48 kHz sample rate, the Nyquist frequency is 24 kHz.

Frequency components above Nyquist cannot be represented directly at that sample rate.

O

Open source

Software whose source code is available under a license permitting specified forms of use, inspection, modification, and redistribution.

Open source does not mean:

No copyright
No license restrictions
Universal model compatibility
Free use for every possible purpose

Always check the project's actual license.

Oversampling

Processing audio internally at a higher sample rate than the system's nominal rate.

Oversampling is commonly used in nonlinear audio processing to reduce audible aliasing.

It increases computational requirements.

P

Pedal capture

A model of a physical effects pedal.

Capture technologies commonly support nonlinear pedals such as:

Boost
Overdrive
Distortion
Fuzz

Support for compressors and other dynamic processors varies by capture architecture.

Playback

Running a completed model with new audio.

For neural models this is inference.

Playback requirements can differ substantially from training requirements.

See NAM Playback.

Preamp capture

A model representing a preamplifier or the preamp section of an amplifier.

A preamp capture normally excludes:

Power amplifier
Speaker
Cabinet
Microphone

unless those elements are deliberately included elsewhere in the measured path.

Profile

Generic term sometimes used for a captured representation of equipment.

PROFILE is also Kemper's specific term for the result of the Kemper Profiling process.

See Kemper Profiling.

Profiling

A term for measuring equipment and creating a digital representation of its behavior.

The term is particularly associated with Kemper, although it is also used generically.

Proteus

An open-source neural-modeling project developed by GuitarML.

Proteus supports snapshot and conditioned models.

See Open Capture Technologies.

Proxy

Line 6's capture technology introduced for Helix Stadium.

The resulting models are called Clones.

See Line 6 Proxy.

Q

Quad Cortex

Neural DSP's Cortex hardware processor providing amplifier/effects modeling and Neural Capture functionality.

It supports creation and playback of Neural Captures.

R

Reactive load

A load device designed to present an amplifier with an electrical load whose impedance varies with frequency in a manner intended to resemble a loudspeaker system.

Reactive load boxes are widely used for direct amplifier recording and capture.

Reamp box

A device used to interface a line-level playback source with equipment designed for an instrument-level input.

Typical applications include sending a recorded signal from an audio interface into:

Guitar amplifier
Pedal
Other instrument-level equipment

A reamp box may provide:

Level attenuation
Balanced-to-unbalanced conversion
Ground isolation
Appropriate source impedance

See Reamping for Capture.

Reamping

Sending previously recorded audio back through physical audio equipment.

In capture workflows, reamping allows a standardized training signal to be sent repeatedly through different equipment or settings.

See Reamping for Capture.

Reference level

A defined signal level used as a known measurement point.

Reference levels are important when relating digital levels to analog voltage.

Return

The signal coming back from the device under test into the capture or recording system.

For example:

interface output → amplifier → load box → interface return

Rig

A collection of equipment forming a usable signal chain.

Depending on context, a rig may include:

Pedals
Amplifier
Cabinet
Microphone
Effects

In Kemper terminology, Rig also has a specific platform meaning.

Rig Exchange

Kemper's online service for sharing PROFILER Rigs.

RMS

Root Mean Square.

A method of describing signal magnitude.

For a sine wave, RMS voltage is commonly used when calculating dBu during calibration.

RTNeural

An open-source neural-network inference engine optimized for real-time audio.

RTNeural is used by projects including AIDA-X and GuitarML software.

RTNeural itself is not a capture-model format.

S

Sample rate

The number of digital audio samples processed per second.

Common rates include:

44.1 kHz
48 kHz
88.2 kHz
96 kHz

Capture/training systems may require a specific sample rate.

Send

The signal leaving the capture or recording system and traveling toward the device under test.

For example:

interface send → reamp box → amplifier

Signal chain

The ordered path through which an audio signal travels.

The signal chain determines which equipment contributes to a capture.

Snapshot

See Snapshot model.

Snapshot model

A model representing equipment at a particular configuration.

For example:

amplifier gain = 5

and:

amplifier gain = 8

may require separate snapshot models.

Snapshot modeling contrasts with conditioned modeling, in which one model can respond to a control parameter.

Stimulus

See Capture signal.

Studio capture

A capture of studio equipment such as:

Preamp
Saturation device
Channel strip
Compressor

Whether a particular device can be captured accurately depends on the capabilities of the capture architecture.

System identification

The engineering process of constructing a mathematical model of a system from observed input and output behavior.

Capture and profiling technologies can be viewed as forms of nonlinear system identification.

T

Test signal

See Capture signal.

Tone Model

IK Multimedia's term for a model created for the TONEX ecosystem using AI Machine Modeling.

A Tone Model is not a NAM model.

ToneNET

IK Multimedia's online service for discovering and sharing TONEX Tone Models and related content.

Training

The computational process that adjusts a model's parameters so its predicted output approximates the measured output of the physical device.

Conceptually:

known input + measured output → training → model

Training should be distinguished from the physical recording or capture stage.

Training data

Input/output measurements used to train a model.

For capture technology this commonly consists of:

Known training input
Recorded response of the physical equipment

Training error

A numerical measurement of the difference between model predictions and the target training data.

Lower training error can indicate a closer mathematical fit, but it is not a complete measurement of perceived audio quality.

Training signal

See Capture signal.

V

V1

The original Neural DSP Neural Capture generation.

V1 training occurs locally on compatible Cortex hardware.

See Neural Capture.

V2

The second-generation Neural DSP Neural Capture technology.

V2 uses higher-resolution cloud training and is designed to improve reproduction of difficult dynamic targets including fuzzes and compressors.

See Neural Capture.

VST3

A plugin format widely used by digital audio workstations.

Several capture-model players are available as VST3 plugins.

W

WaveNet

A convolutional neural-network architecture originally developed for audio generation.

WaveNet-style architectures have also been adapted for nonlinear audio-system modeling.

Some GuitarML projects have used WaveNet-based architectures.