Open Capture Technologies
Open Capture Technologies
Open capture technologies are open-source projects that use machine learning or related techniques to reproduce the behavior of amplifiers, pedals, preamps, and other audio equipment.
Neural Amp Modeler is currently one of the best-known open capture ecosystems, but it is not the only open-source project in this field.
Other projects have explored different:
- Neural-network architectures
- Training methods
- Model formats
- Playback engines
- Plugin formats
- Embedded hardware
- Conditioned models
- Snapshot models
Important examples include AIDA-X and projects developed by GuitarML, including Proteus, NeuralPi, and earlier SmartGuitarAmp/SmartPedal work.
These technologies are related conceptually but should not be assumed to be model-compatible.
Open source does not mean one format
There is no universal open-source capture-model format.
An open-source project may define its own:
- Model architecture
- File format
- Training procedure
- Sample rate
- Input normalization
- Runtime engine
- Metadata
- Playback requirements
Consequently:
open-source model ≠ NAM model
and:
neural amp model ≠ .nam file
A model created for one project generally cannot be loaded directly by another unless explicit compatibility has been implemented.
Why open capture technologies matter
Open implementations allow researchers, developers, musicians, and hardware manufacturers to examine and modify the underlying technology.
Depending on the project's license and architecture, developers may be able to:
- Study the training process
- Study the playback implementation
- Build alternative players
- Create hardware implementations
- Modify the neural architecture
- Develop new training tools
- Experiment with model conditioning
- Improve performance
- Create independent model libraries
This differs from proprietary capture systems whose internal model representation and playback technology are controlled by one manufacturer.
See Capture Technologies.
AIDA-X
AIDA-X is an open-source neural Amp Model player developed by Aida DSP.
Its primary purpose is high-fidelity modeling of guitar amplifiers, although models can represent larger signal chains containing combinations of:
- Amplifier
- Cabinet
- Distortion
- Overdrive
- Fuzz
- Boost
- EQ
AIDA-X uses neural-network models trained from recorded input/output examples.
The project provides both desktop/plugin playback and implementations intended for embedded hardware.
AIDA-X model playback
AIDA-X uses RTNeural as its neural-network inference engine.
The desktop implementation supports:
- CLAP
- LV2
- VST2
- VST3
- Standalone operation
Installers or binaries are available for Windows, macOS, and Linux.
The project also maintains a headless LV2 implementation intended for embedded environments.
Examples include systems based on:
- MOD hardware
- Raspberry Pi
- Portenta X8
- Aida DSP OS
This makes AIDA-X notable not only as a desktop neural amplifier project but also as an open platform for embedded neural-model playback.
AIDA-X cabinet processing
AIDA-X includes cabinet impulse-response processing after the Amp Model.
Users can therefore combine a direct amplifier model with a cabinet IR within the player.
Supported IR audio formats include WAV and FLAC.
This follows the common modular signal-chain approach:
neural amplifier model → cabinet IR
See Cabinets and IRs.
Training AIDA-X models
The AIDA-X project provides a model-training workflow.
As with other capture systems, the general process involves:
- Obtain or generate the required training input.
- Send the signal through the equipment being modeled.
- Record the resulting output.
- Train the neural model from the corresponding input and output.
- Load the finished model into AIDA-X.
- Compare the model with the original equipment.
The exact current training procedure should be obtained from the project's current trainer and documentation because open-source training tools can evolve independently of the playback application.
AIDA-X and NAM
AIDA-X and Neural Amp Modeler are separate projects.
Although both use neural networks to reproduce audio equipment, an AIDA-X model should not automatically be assumed to be a .nam model.
The projects have their own:
- Model architectures
- Training systems
- Playback implementations
- Model formats
The existence of RTNeural in one project and neural-network processing in another does not make their models interchangeable.
Compatibility should always be verified explicitly.
GuitarML
GuitarML is an open-source collection of machine-learning projects focused on guitar amplifiers, pedals, and related audio processing.
GuitarML has produced several experimental and practical neural-modeling systems, including:
- Proteus
- NeuralPi
- SmartGuitarAmp
- SmartPedal
- SmartAmp-related projects
- Training utilities and model libraries
These projects have explored multiple neural-network architectures rather than defining one universal GuitarML model format.
GuitarML Proteus
Proteus is an open-source GuitarML plugin for capturing and playing models of amplifiers, pedals, plugins, and complete rigs.
The project describes its purpose as providing functionality conceptually similar to systems such as Kemper, Quad Cortex, and TONEX through a free and open-source plugin.
Proteus can create:
- Snapshot models
- Conditioned knob models
This distinction makes Proteus particularly interesting within the history of open capture technology.
Proteus snapshot models
A Proteus snapshot represents equipment at a particular operating point.
Examples include:
- Amplifier at one gain setting
- Overdrive at one drive setting
- Pedal/amplifier combination
- Complete gain-based rig
This is conceptually similar to snapshot capture in several other systems.
The physical controls are configured before the training recording is made.
Proteus conditioned models
Proteus can also capture the behavior of a variable control.
The project refers to these as Knob models.
Instead of learning only one fixed operating point, training data is recorded at multiple control positions.
The resulting model can then respond to a parameter representing that control.
This is an important distinction between:
snapshot modeling
and:
conditioned modeling
A snapshot learns:
audio input → audio output
at a particular equipment configuration.
A conditioned model can conceptually learn:
audio input + control value → audio output
allowing one neural model to represent part of the equipment's behavior across a control range.
Proteus neural architecture
Proteus uses an LSTM neural network.
The current project documentation specifies an LSTM hidden size of 40 for compatible Proteus models.
The real-time neural-network processing is performed using RTNeural.
Proteus model files use JSON.
These files are specific to the Proteus architecture and should not be confused with NAM files merely because both formats can use JSON internally.
Capturing with Proteus
The Proteus Capture Utility provides the input audio and training workflow required to create models.
A typical capture process is:
- Play the Proteus capture WAV through the equipment.
- Record the equipment output.
- Export the recording.
- Upload the recording to the provided Google Colab training workflow.
- Train the model.
- Download the resulting JSON model.
- Load the model into Proteus.
The project recommends using a reamp device where appropriate and a load box for direct amplifier captures.
See Reamping for Capture.
Proteus capture file requirements
The current Proteus documentation specifies capture recordings as:
- WAV
- Mono
- 44.1 kHz
- PCM16 or FP32
Timing alignment is particularly important.
The project warns that latency mismatch between the training input and recorded output can prevent successful training.
An initial click in the supplied capture signal can be used as an alignment reference.
This illustrates a principle shared by many capture technologies:
The training system must know which output samples correspond to which input samples.
Proteus model targets
Proteus is intended primarily for nonlinear equipment such as:
- Amplifiers
- Preamps
- Distortion pedals
- Overdrive pedals
- Boost pedals
The project explicitly advises against attempting to capture time-based effects such as:
- Reverb
- Delay
- Flanger
- Phaser
A larger gain-based signal chain can also be captured as one model.
For example:
overdrive → amplifier
can be learned as one combined system.
Proteus direct amplifier models
A direct amplifier model excludes the physical cabinet and microphone.
Proteus recommends combining these models with an impulse response during playback.
The signal chain becomes:
Proteus amp model → cabinet IR
Proteus includes a default IR for convenience, although the project recommends using a dedicated IR solution where appropriate.
Proteus Amp+Cab models
An amplifier can also be captured using a microphone.
In that case, the resulting model includes the measured:
- Amplifier
- Speaker
- Cabinet
- Microphone
- Microphone position
An additional cabinet IR is normally unnecessary unless deliberately used for creative processing.
See Cabinets and IRs.
Proteus plugin formats
Proteus is available in common desktop plugin formats.
Current project releases provide formats including:
- VST3
- AU
- AAX
- LV2
Platform availability varies by operating system.
Because the project is open source, the code can also be compiled independently.
GuitarML NeuralPi
NeuralPi is another GuitarML neural-modeling project.
It was developed partly around running neural amplifier models on compact and embedded systems, including Raspberry Pi-related applications.
NeuralPi supports both:
- Snapshot models
- Conditioned models
Later versions added the ability to load models conditioned on a Gain/Drive parameter.
This represents another exploration of variable-control neural modeling.
GuitarML model compatibility
Even models created by different GuitarML projects are not necessarily interchangeable.
The GuitarML ToneLibrary identifies different neural architectures for its projects.
Examples include:
Proteus
Stateful LSTM followed by a dense layer.
NeuralPi
Stateful LSTM followed by a dense layer.
SmartAmpPro
Convolutional layers followed by a stateless LSTM and dense layer.
SmartPedal / SmartAmp
WaveNet.
Because the architectures differ, models should be loaded only into the software for which they were created unless explicit conversion or compatibility exists.
GuitarML ToneLibrary
GuitarML maintains an open ToneLibrary containing model files for its projects.
The library organizes models according to the intended player.
This is necessary because a JSON model for one GuitarML architecture may not work with another GuitarML application.
File extension alone is therefore not sufficient to determine compatibility.
SmartGuitarAmp
SmartGuitarAmp was an earlier GuitarML project using a WaveNet neural network to reproduce amplifier behavior.
The project helped demonstrate practical real-time neural guitar processing in an open-source plugin.
Later GuitarML projects explored other architectures and more flexible model-loading and capture workflows.
SmartGuitarAmp is historically important as part of the development path that led to the broader GuitarML ecosystem.
RTNeural
RTNeural is a high-performance neural-network inference engine designed for real-time audio applications.
It is used by projects including:
- AIDA-X
- GuitarML Proteus
RTNeural is not itself an amplifier capture format.
It provides infrastructure for efficiently executing supported neural-network architectures.
Different applications can use RTNeural while still having completely incompatible model formats and architectures.
This distinction is important.
A common runtime technology does not imply a common model standard.
Snapshot versus conditioned modeling
Open projects have been particularly useful for experimenting with the distinction between snapshot and conditioned models.
A snapshot model represents one equipment configuration.
For example:
Amp gain = 6
A separate capture might represent:
Amp gain = 8
A conditioned model can instead accept a control parameter as part of the neural-network input.
Conceptually:
audio + gain setting → modeled audio
This allows one trained model to reproduce behavior across multiple settings.
Conditioned modeling can potentially reduce the need for large collections of individual snapshots.
However, training and accurately reproducing a variable control can be more complex than modeling one fixed operating point.
Capture technology is broader than products
Commercial capture systems are often discussed as products:
Open-source capture technology is better understood as an evolving collection of:
- Research
- Code
- Training systems
- Runtime libraries
- File formats
- Plugins
- Hardware implementations
A research project can influence later products even if the original software itself never becomes a widely used commercial platform.
Research origins
Modern neural amplifier modeling builds on academic research into real-time nonlinear audio-system modeling.
Important approaches have included:
- Recurrent neural networks
- LSTM networks
- WaveNet-style convolutional networks
- Conditioned neural networks
- Other temporal neural architectures
Open-source projects have made many of these techniques accessible to musicians and developers without requiring proprietary hardware.
Model files are not audio recordings
A neural model file is not simply the captured output recording.
The recording is training data.
Training analyzes the relationship between:
known input
and:
recorded output
and produces a neural-network model.
The model contains parameters that allow a compatible player to predict the equipment's response to new audio.
Therefore:
capture recording ≠ trained model
Preserving both can be valuable.
Preserving training data
Where the training system permits it, creators should preserve:
- Original training input
- Original recorded output
- Equipment settings
- Interface settings
- Calibration information
- Training software/version
- Finished model
- Notes
The original measurement may remain useful even if the model architecture or training software changes later.
A new neural architecture can potentially be trained from old measurement data if the new system accepts compatible training material.
Open formats and longevity
One potential advantage of open capture technologies is long-term inspectability.
If the:
- Model specification
- Playback code
- Training code
remain publicly available, models may remain usable even if the original developer stops maintaining a particular application.
This does not guarantee permanent compatibility.
Open-source projects can still change:
- Architecture
- File format
- Dependencies
- Operating-system support
- Build systems
Archiving the relevant source code and documentation can therefore be as important as preserving model files.
Open source and commercial use
"Open source" does not necessarily mean that code can be incorporated into any commercial product without restrictions.
Each project has a software license.
Examples include licenses such as:
- GPL
- MIT
- BSD
- Other open-source licenses
The rights and obligations differ substantially.
Developers considering integration into hardware or commercial software should examine the actual license of every relevant component.
Do not assume that freely downloadable source code is equivalent to unrestricted commercial licensing.
Comparing open capture systems
Useful comparison criteria include:
- Modeling accuracy
- Neural architecture
- Snapshot support
- Conditioned-model support
- Training time
- Training hardware
- CPU requirements
- Latency
- Model file format
- Sample rate
- Calibration
- Metadata
- Cabinet handling
- Embedded support
- Plugin support
- Licensing
- Model sharing
- Development activity
No single criterion determines whether a technology is useful.
A small neural architecture may be ideal for embedded hardware even if a larger model achieves lower numerical error on a desktop computer.
Interoperability
Open source makes interoperability possible, but it does not make interoperability automatic.
For example:
AIDA-X model → Proteus
is not automatically supported.
Proteus model → NAM
is not automatically supported.
NAM model → AIDA-X
is not automatically supported.
A converter or compatible playback implementation must understand the source architecture and represent it correctly.
Simply renaming a file extension does not convert a neural model.
Why multiple architectures exist
Different neural architectures represent different engineering tradeoffs.
Developers may optimize for:
- Accuracy
- CPU usage
- Memory
- Training speed
- Embedded hardware
- Parameter conditioning
- Development simplicity
- Real-time performance
The existence of multiple open projects is therefore useful.
They provide different approaches to the same broad problem rather than requiring one implementation to satisfy every use case.
Relationship to NAM
Neural Amp Modeler is part of this broader open-source neural audio field.
NAM has developed its own:
- Model specification
- Training system
- Neural architectures
- Real-time core
- Plugin/player ecosystem
- Hardware integrations
- Model-sharing ecosystem
Other open projects should not merely be described as NAM alternatives.
They represent independent lines of development that have contributed to the larger field of neural audio modeling.
Documentation
When sharing a model from an open capture system, identify at minimum:
- Project/platform
- Compatible player
- Model architecture where relevant
- Equipment
- Capture type
- Equipment settings
- Cabinet status
- Sample rate where relevant
- Training method/version
- Creator
- Notes
This is particularly important when several systems use generic extensions such as .json.
A filename alone may provide no reliable indication of model compatibility.
Troubleshooting
If an open-source model will not load, first determine:
- Which project created the model.
- Which architecture it uses.
- Which player it was intended for.
- Whether the player version supports that architecture.
- Whether the model file is complete and valid.
If the model loads but sounds incorrect, check:
- Input level
- Sample rate
- Cabinet processing
- IR configuration
- Calibration
- Output level
- Additional processing
- Training quality
Do not assume that two neural-model players will produce equivalent results merely because both are open source.
Official resources
See also
- Capture Technologies
- Neural Amp Modeler
- NAM Model File Format
- Training a NAM Model
- Capture Types
- Creating a Capture
- Capture Signal Chain
- Reamping for Capture
- Cabinets and IRs
- Troubleshooting Captures
Discuss AIDA-X, GuitarML, Proteus, and other open capture technologies on NAMFORUM