NAMFORUM Sysop: Created page

2026-09-01T06:03:27Z

Created page

New page

= NAM Playback =

'''NAM playback''' is the real-time execution of a trained [[Neural Amp Modeler]] model.

A trained ''.nam'' file contains the neural-network model, but the file does not process audio by itself. It must be loaded into compatible software or hardware that implements the model architecture.

NAM models can be played through:

* Audio plugins
* Standalone computer applications
* Dedicated hardware
* Multi-effects processors
* Embedded systems
* Third-party applications incorporating NAM technology

Because NAM's real-time DSP implementation is open source, the same model format can be implemented by many independent developers and manufacturers.

== Playback signal flow ==

A basic NAM playback chain is:

'''Instrument → audio interface → NAM player → output'''

For a direct amplifier model, the chain may instead be:

'''Instrument → NAM model → cabinet IR/model → output'''

For a model that already includes a cabinet and microphone:

'''Instrument → NAM model → output'''

The correct configuration depends on what was included in the original capture.

See:

* [[Capture Types]]
* [[Cabinets and IRs]]

== NeuralAmpModelerCore ==

'''NeuralAmpModelerCore''' is the open-source C++ DSP library maintained by the NAM project for playing NAM models in real time.

It provides the underlying model-execution functionality that developers can integrate into:

* Plugins
* Applications
* Hardware
* Embedded systems

This separation between the model format and playback implementation is fundamental to NAM.

A ''.nam'' model is not inherently tied to one particular plugin, computer, pedal, or manufacturer.

== Official NAM playback software ==

The NAM project provides official software for playing NAM models.

The current user-facing NAM player is '''Gateway'''.

Gateway loads and plays NAM snapshot models and is available for Windows and macOS.

The NAM project also maintains the underlying plugin code used to integrate NeuralAmpModelerCore into desktop audio software.

Historically, the official NeuralAmpModelerPlugin has provided:

* VST3 plugin support
* Audio Unit support on macOS
* Standalone desktop operation

The exact available formats and applications may change as NAM software develops.

Consult the current NAM website for downloads.

== Loading a model ==

The basic playback process is:

# Obtain a compatible ''.nam'' model.
# Open a NAM-compatible player.
# Load the model.
# Determine whether the model contains a cabinet.
# Add cabinet processing if required.
# Set appropriate input and output levels.
# Play through the model.

Loading the model is only one part of reproducing the original captured sound.

The surrounding signal chain also matters.

== Input level ==

Input level is particularly important with NAM because most NAM models represent nonlinear equipment.

Increasing the signal entering the model can change:

* Distortion
* Saturation
* Compression
* Attack
* Sustain
* Dynamic response

This is different from simply changing the volume after the model.

If a model was captured with calibration metadata, compatible playback systems can use that information to help reproduce the electrical level relationship of the original equipment.

See [[Gain Staging and Calibration]].

== Calibration-aware playback ==

NAM model files can contain:

'''input_level_dbu'''

and:

'''output_level_dbu'''

metadata.

These describe the relationship between digital levels and the analog levels used during capture.

A calibration-aware playback system can use this information to help determine how strongly the virtual equipment should be driven and how its output level relates to the original device.

This can make a model more portable between systems with different audio-interface calibration.

Calibration does not simply determine loudness.

For nonlinear equipment, incorrect input calibration can change the actual modeled behavior.

See [[NAM Model File Format]].

== Models without calibration metadata ==

Many NAM models do not contain calibration metadata.

They can still be used.

In that case, the player cannot automatically know the precise analog level relationship that existed during the original capture.

The user may therefore need to adjust the input level by ear or according to information supplied by the model creator.

This is one reason good capture documentation remains important.

== Direct amplifier models ==

A direct amplifier NAM normally excludes the physical speaker cabinet and microphone.

When played through studio monitors, headphones, PA speakers, or another full-range system, it will normally require:

* A cabinet IR
* A cabinet model
* Another suitable speaker simulation

A typical chain is:

'''NAM amplifier model → cabinet IR → full-range output'''

Without cabinet processing, a direct guitar-amplifier model may sound unusually bright, harsh, or aggressive.

That does not necessarily indicate a defective model.

== Models containing cabinets ==

A model may already contain the response of:

* Amplifier
* Speaker
* Cabinet
* Microphone
* Microphone position
* Microphone preamp

In this case, another conventional cabinet IR is normally unnecessary.

Adding one may result in unintended double cabinet processing.

Before adding an IR, determine what the NAM actually contains.

See [[Cabinets and IRs]].

== Playing through a physical guitar cabinet ==

A NAM amplifier model can also be used with a traditional guitar cabinet.

A typical chain is:

'''Instrument → NAM amplifier model → suitable power amplifier → guitar cabinet'''

In this configuration, a cabinet IR is often disabled because the physical guitar cabinet supplies the speaker and cabinet coloration.

The power-amplification system must be suitable for the cabinet and intended operating level.

== Full-range playback ==

NAM models containing cabinet processing are commonly reproduced through:

* Studio monitors
* Headphones
* PA systems
* FRFR systems
* Full-range powered speakers

These systems are intended to reproduce the complete modeled signal rather than add the strong coloration of a traditional guitar loudspeaker.

No real playback system is perfectly neutral, so the monitoring environment still influences what is heard.

== Native NAM playback ==

A system providing '''native NAM playback''' executes the NAM model itself using a compatible implementation of its neural-network architecture.

NeuralAmpModelerCore is the reference open-source implementation used by many developers.

In this type of system:

'''.nam model → NAM-compatible DSP implementation → audio'''

The objective is to execute the model represented by the NAM file directly.

== Converted NAM playback ==

Some products allow a ''.nam'' file to be imported but do not execute the NAM neural network directly.

Instead, the NAM model is analyzed and converted or recaptured into another model architecture that the device can execute.

The process is conceptually:

'''.nam model → conversion → proprietary model → audio'''

This can be extremely useful, particularly on hardware that lacks the processing resources required for native NAM execution.

However, it is technically different from running the original NAM model.

The NAM project has described this distinction using an analogy to lossless and lossy audio conversion.

A converted model attempts to reproduce the NAM but may not produce mathematically identical processing.

== Why the distinction matters ==

Both native and converted systems can provide useful access to the NAM ecosystem.

The distinction becomes important when discussing:

* Model accuracy
* Architecture compatibility
* CPU requirements
* A1 versus A2 support
* Troubleshooting
* Comparisons between playback devices

A product that can '''import a .nam file''' should not automatically be assumed to execute that NAM model natively.

When this distinction matters, consult the manufacturer's technical documentation.

== A1 playback ==

The original standard NAM WaveNet architectures are now referred to collectively as '''A1'''.

These include:

* A1-standard
* A1-lite
* A1-feather
* A1-nano

Large libraries of A1 models exist, and current NAM implementations continue to support them.

A1 remains an important part of the NAM ecosystem.

== A2 playback ==

[[NAM A2 Architecture|A2]] became NAM's new standard architecture on June 2, 2026.

A2-capable playback requires an implementation updated to understand the newer architecture.

The NAM project identified the following versions at the A2 release as supporting the architecture:

* neural-amp-modeler v0.13.0
* NeuralAmpModelerCore v0.5.2
* NeuralAmpModelerPlugin v0.7.14

Later versions should normally retain support.

Older NAM software or hardware that predates A2 support may need an update before it can load A2 models.

== Backward compatibility ==

Current NAM implementations are designed to retain support for existing A1 models.

Therefore:

'''A2-capable player → A2 model: supported'''

'''A2-capable player → A1 model: supported'''

An older A1-only player, however, cannot automatically be expected to understand A2.

See [[NAM A2 Architecture]].

== A2 slimmable playback ==

A2 incorporates slimmable architecture capability.

This allows an A2 model to be executed at different computational levels where the playback implementation exposes that capability.

A player may therefore offer a higher-accuracy mode and a reduced-computation mode using the same A2 model.

The exact user controls depend on the playback product.

This can be particularly useful in hardware where DSP resources must be shared among multiple effects.

== Hardware playback ==

NAM has been implemented in a growing range of dedicated and multi-effects hardware.

Hardware implementations can differ substantially.

Possible approaches include:

* Native NeuralAmpModelerCore execution
* Independent compatible NAM implementation
* Conversion into another neural model format

Hardware may also differ in:

* Supported NAM architectures
* Maximum number of simultaneous models
* Sample rate
* Cabinet processing
* Model storage
* Calibration support
* CPU allocation
* Model-management workflow

Therefore, "NAM compatible" does not necessarily describe one identical set of capabilities.

== HeadRush example ==

In August 2026, Firmware 5.1 added native A1 and A2 NAM playback to the HeadRush:

* Prime
* Core
* Flex Prime

NAM models can be transferred from a computer or accessed through TONE3000 integration.

The implementation also demonstrates A2's slimmable capability through a Lite Mode that reduces processor usage.

This is one example of a commercial device executing current NAM architectures directly.

== Converted hardware implementations ==

Other manufacturers have used conversion approaches to make NAM models accessible on processors that run different internal architectures.

Examples documented by the NAM project have included products from manufacturers such as:

* Hotone
* Valeton
* Sonicake
* Sonulab

In these systems, importing a NAM can involve creating a device-specific representation of the source model.

This should not automatically be interpreted as inferior or undesirable.

It represents a different engineering tradeoff.

The important point is to describe the implementation accurately.

== Software playback ==

NAM can also be incorporated into third-party software.

Because the real-time DSP implementation is available under an open-source license, developers can build NAM support into:

* Amp suites
* Effects processors
* DAWs
* Standalone applications
* Experimental software
* Research systems

Some applications may combine NAM with:

* Cabinet IRs
* Effects
* Routing
* Preset management
* Model libraries
* Other modeling technologies

The resulting user experience can therefore vary considerably even when the same NAM model is being executed.

== CPU usage ==

Neural-network models require computation for every audio sample processed.

CPU requirements depend on factors including:

* Model architecture
* Architecture configuration
* Sample rate
* Number of simultaneous models
* Host implementation
* Processor architecture
* Optimization
* A2 operating mode where applicable

A model that runs easily on a desktop computer may represent a significant workload on an embedded processor.

A2 was developed partly to improve this relationship between accuracy and computational cost.

== Latency ==

Real-time NAM playback is part of a larger digital audio system.

Total playing latency can include:

* Audio-interface input buffering
* A/D conversion
* Host processing buffer
* NAM processing
* Additional plugins
* D/A conversion
* Digital hardware routing

The NAM model itself is therefore only one component of total system latency.

For comfortable live playing, configure the complete system appropriately rather than evaluating buffer settings in isolation.

== Sample rate ==

A NAM model can contain sample-rate information.

If the sample-rate field is absent, the current NAM file specification states that implementations should generally assume 48 kHz.

Playback implementations are responsible for handling the relationship between the model's expected sample rate and the host system appropriately.

See [[NAM Model File Format]].

== Comparing NAM players ==

When comparing two NAM playback systems, use the same model and control as many variables as possible.

Check:

* Native versus converted playback
* A1/A2 architecture
* Input calibration
* Input level
* Output level
* Cabinet processing
* IR
* Sample rate
* Additional EQ
* Effects
* Monitoring system

Otherwise, an apparent difference between NAM implementations may actually come from the surrounding signal chain.

== Model libraries ==

The official NAM user page currently directs users to '''TONE3000''' for sharing and discovering models.

Models can also be distributed independently because a ''.nam'' file is portable.

A model downloaded from any source should be accompanied by enough information to determine:

* What equipment it represents
* Whether a cabinet is included
* Model architecture
* Relevant equipment settings
* Calibration where available

The playback system cannot infer every aspect of the original physical rig merely from the sound of the model.

== Troubleshooting playback ==

If a NAM model will not load, check:

# Is the file valid?
# Is it actually a .nam model?
# Is it A1 or A2?
# Does the player support that architecture?
# Is the player or firmware current?
# Does the device use native NAM or require conversion?

If the model loads but sounds wrong, check:

# Input level
# Calibration
# Cabinet/IR configuration
# Output level
# Additional processing
# Playback system

Do not assume immediately that the model itself is defective.

See [[Troubleshooting Captures]].

== The importance of portability ==

One of NAM's defining characteristics is that model creation and model playback are separate.

A capture can be:

* Recorded by one person
* Trained using one service
* Shared through another service
* Downloaded by another user
* Played in software
* Played in compatible hardware

No single playback product is required for the model to exist.

This separation between the '''model''' and the '''player''' is one of the principal reasons NAM can function as an ecosystem rather than merely as a feature of one manufacturer's product.

== Official resources ==

* [https://www.neuralampmodeler.com/users NAM user and playback resources]
* [https://www.neuralampmodeler.com/the-code NAM open-source repositories]
* [https://www.neuralampmodeler.com/post/a2-is-released A2 release announcement]
* [https://www.neuralampmodeler.com/ Neural Amp Modeler official website]

== See also ==

* [[Neural Amp Modeler]]
* [[NAM Model File Format]]
* [[Training a NAM Model]]
* [[NAM A2 Architecture]]
* [[Gain Staging and Calibration]]
* [[Capture Types]]
* [[Cabinets and IRs]]
* [[Troubleshooting Captures]]

'''[https://namforum.com/ Discuss NAM players, hardware, and model compatibility on NAMFORUM]'''

[[Category:Neural Amp Modeler]]
[[Category:NAM playback]]
[[Category:Capture technology]]

NAM Playback - Revision history

NAMFORUM Sysop: Created page