Training a NAM Model
Training a NAM Model
Training a NAM model is the process of using recorded input and output audio to create a Neural Amp Modeler model of an amplifier, pedal, preamp, or other nonlinear audio system.
NAM learns the relationship between a known input signal and the signal produced after that input passes through the equipment being modeled.
The basic process is:
NAM input file → device under test → recorded output file → NAM trainer → .nam model
Creating a good model therefore begins before training. The quality and accuracy of the recorded data are fundamental to the result.
Before training
Before creating training data, decide exactly what the model should represent.
The device under test may be:
- An amplifier
- A pedal
- A preamp
- A direct amplifier signal
- An amplifier and cabinet
- Multiple devices in a chain
Configure the equipment and document its settings.
See:
Obtain the official input file
The simplified NAM training workflow provides a standardized input audio file.
The current local GUI trainer includes a Download input file button that links to the input audio used for training.
Use the input file intended for the version and workflow of the NAM trainer being used.
Do not casually modify the training file.
Changes to its:
- Level
- Length
- Sample rate
- Bit depth
- Timing
- Content
can interfere with training or alter the relationship between the expected input and recorded output.
Reamp the input through the equipment
The NAM input file is sent through the equipment being modeled.
A typical path is:
NAM input.wav → audio interface output → reamp/capture path → device under test → return path → audio interface input
The resulting signal is recorded as the output file.
See Reamping for Capture.
Depending on the equipment, the return path might be:
Pedal → interface input
or:
Amplifier → suitable load/direct capture device → interface input
or:
Amplifier → cabinet → microphone → preamp/interface input
The recorded output contains the behavior NAM will attempt to learn.
File format
For the current simplified NAM GUI training workflow, the official input file is:
- 48 kHz
- 24-bit
- WAVE
The rendered output should match those characteristics.
The output recording must also match the length of the supplied input file.
These requirements should be checked against the current NAM documentation whenever the trainer or workflow changes.
Record the return cleanly
Unless intentionally part of the device being modeled, avoid adding processing to the recorded return.
Check for:
- DAW plugins
- EQ
- Compression
- Limiting
- Normalization
- Noise reduction
- Cabinet simulation
- Interface DSP
- Automatic gain control
- Master-bus processing
Anything in the measured path can influence the model.
If processing is deliberately part of the target chain, document it.
Avoid clipping
The return recording must not contain unintended clipping.
Check every stage of the capture path rather than relying only on the final DAW meter.
Possible overload points include:
- Interface output
- Reamp device
- Device under test
- Load/direct-output equipment
- Microphone preamp
- Interface input
- A/D converter
- DAW signal path
Intentional distortion produced by the equipment being modeled is different from unintended clipping elsewhere in the recording system.
Send level
The NAM calibration documentation recommends providing the equipment with training examples at least as loud as the levels at which the resulting model is expected to be used.
This is particularly important with nonlinear equipment.
If the training signal never drives the equipment into a region that will later be demanded from the model, the trainer has no measured examples of that behavior from which to learn.
This does not mean that arbitrary overload is desirable.
The objective is to expose the equipment to the useful operating range that the model is expected to reproduce while maintaining a valid capture signal chain.
See Gain Staging and Calibration.
Calibration
NAM supports optional calibration metadata describing the relationship between digital signal level and the analog levels used during capture.
The relevant metadata fields are:
input_level_dbu
and:
output_level_dbu
Providing these values is optional. A NAM model can operate without calibration metadata.
However, calibrated models can allow compatible playback systems to reproduce the gain relationship of the original analog equipment more accurately.
The official NAM calibration procedure involves measuring the analog send level and determining the return-interface level corresponding to digital full scale.
See:
Do not change hardware gain after calibration measurement
Calibration describes the actual hardware configuration used to create the training data.
If output or input gain controls are changed after recording, calibration measurements made using the new settings no longer describe the original capture.
For repeatable work, document and preserve:
- Interface output setting
- Reamp setting
- Interface input gain
- Other variable gain stages
until the relevant measurements are complete.
Preserve the original recording
Keep an untouched copy of the returned training recording.
This is useful if:
- Training must be repeated
- A new architecture becomes available
- Metadata must be corrected
- A problem is discovered
- Another trainer is used
- Training parameters are compared
Do not make destructive edits to the only copy of the original capture data.
Input and output length
The current simplified NAM trainer requires the returned output recording to match the length of the input file.
Problems can arise if:
- Recording starts late
- Recording stops early
- Extra silence is added
- Material is removed
- The DAW adds a tail
- Export boundaries are wrong
If editing is required to meet the trainer's specification, preserve the original recording separately.
Latency alignment
A recording system introduces round-trip latency.
This means the recorded response does not necessarily begin at exactly the same sample position as the original training signal.
The current NAM GUI trainer automatically attempts to align the input and output recordings.
The official NAM input file contains two impulses near its beginning that assist this process.
During training, the GUI displays its detected alignment so the user can inspect whether the response and input appear correctly aligned.
DAWs may also perform latency compensation, and in some configurations that compensation can be incorrect or excessive.
Therefore, do not assume that DAW compensation automatically guarantees correct alignment.
Select the files
In the local GUI trainer, select:
- The NAM input file
- The recorded output file
- The location where the trained model should be saved
Once valid files have been selected, the trainer enables the training process.
The trainer checks the files for problems before training begins.
Batch training
The NAM GUI supports training multiple models in a batch.
Multiple reamped output files can be selected while using the corresponding input training file.
This can be useful when several captures have been made from:
- Different amplifiers
- Different gain settings
- Different pedal settings
- Different channels
- Different signal chains
Batch training can reduce repetitive setup when processing a larger capture session.
Training
Start training using the trainer.
The software analyzes the relationship between the known input signal and the recorded response and optimizes a neural-network model to reproduce that behavior.
Training involves repeatedly adjusting the model parameters to reduce the difference between:
- The actual recorded output
- The output predicted by the model
When training completes, the resulting model is exported as a .nam file.
Training evaluation
At the end of the current GUI training process, NAM displays a comparison between the model's prediction and the recorded response.
This provides useful information about how closely the trained model follows the captured data.
However, a successful training run is not by itself proof that the model is perceptually accurate.
The model should still be evaluated by listening.
Listen to the model
Load the resulting .nam file into a compatible NAM player.
Compare it with the original equipment under controlled conditions.
Where practical:
- Use the same source performance.
- Match playback levels.
- Use equivalent cabinet processing.
- Switch between original and model quickly.
- Test different playing dynamics.
- Test more than one instrument or source where appropriate.
Listen for differences in:
- Gain
- Distortion
- Dynamics
- Attack
- Compression
- Low-frequency response
- High-frequency response
- Sustain
- Noise
- Transient behavior
A model can achieve a technically successful training result while still revealing audible differences in practical use.
Model architecture
NAM has supported multiple neural-network architectures.
The architecture affects factors including:
- Computational requirements
- Model size
- Training behavior
- Playback compatibility
- Modeling performance
Current simplified NAM training workflows use the newer A2 generation for snapshot modeling.
Older A1 models remain widely used.
See NAM A2 Architecture.
Model metadata
When training a model for distribution, provide useful metadata where the trainer supports it.
This can include:
- Model name
- Creator
- Equipment manufacturer
- Equipment model
- Equipment type
- Tone type
- Input calibration
- Output calibration
Metadata makes the model substantially more useful after it has been separated from the original capture session.
Naming models
Use filenames and model names that remain understandable outside the original session.
Useful identifiers may include:
- Manufacturer
- Equipment model
- Channel
- Gain setting
- Cabinet status
- Significant switches or settings
Avoid relying on names such as:
test1.nam
or:
good-one-final2.nam
for models intended to become part of a permanent collection.
Retraining
One advantage of preserving the original input and output recordings is that the same capture data may be trained again.
Reasons for retraining can include:
- New NAM architecture
- Updated trainer
- Different training settings
- Improved training software
- Comparison of architectures
- Corrected metadata
The physical equipment does not necessarily have to be recaptured if the original measurement data remains suitable for the new training method.
Troubleshooting training
If the trainer rejects the files or the resulting model behaves incorrectly, check:
- Correct input file
- Correct output file
- Sample rate
- Bit depth
- File length
- Clipping
- Latency alignment
- Capture routing
- Calibration
- Unwanted processing
- Software version
Do not randomly modify the recording simply to make training complete.
Identify which requirement is not being met.
Reproducibility
For serious capture work, preserve enough information to reproduce the model.
Useful records include:
- Original NAM input file
- Original recorded output
- Device settings
- Complete capture signal chain
- Interface
- Reamp device
- Calibration
- Trainer version
- Model architecture
- Metadata
- Final .nam file
The training file and resulting model are only part of the experiment.
Good documentation allows the result to be understood later.
Official resources
- NAM local GUI training documentation
- NAM calibration documentation
- Neural Amp Modeler official website