Introducing Neural Amp Modeler (NAM) Architecture 2 (A2)

Today we're launching Architecture 2 (A2), the next generation of Neural Amp Modeler (NAM) built by TONE3000, in partnership with NAM creator Steve Atkinson.

A2 lets anyone create hyper-accurate neural captures of analog gear, including amps, guitar and bass pedals, outboard gear, and full signal chains.

It sounds better than any modeler that has come before it, runs on hardware as small as a $3 chip, and is fully open source.

A2 marks a turning point for the industry in three ways:

• Incredibly accurate sound. A2 beats Neural DSP V2, IK Multimedia ToneX, and Line 6 Proxy by a wide margin in both quantitative and blind listening tests (shown below).
• Works anywhere. A2 runs in DAWs and even budget multi-fx pedals, with sound quality better than Quad Cortex at 50% CPU on a $3 ARM Cortex-M7 600MHz chip.
• Fully open source. Any hardware or software maker can support A2 in their product with code that's freely available online.

A2 makes TONE3000 and Neural Amp Modeler possible on mass-market hardware.

Blackstar, Lava Music, Darkglass, HeadRush, Chaos Audio and Dimehead are all supporting A2, with dozens of major companies announcing support later this year.

"We’re excited to partner with Tone3000 and Steve Atkinson to bring native NAM support to the HeadRush platform. This Summer, Prime, Core, and Flex Prime users will be able to load NAM captures directly to their rigs, with no lossy conversion. Our touchscreen and integrated Wi-Fi will also provide onboard access to TONE3000’s growing library, making it easy to discover, download, and start playing with NAM captures directly on your HeadRush pedal.”
– Walter Skorupski, Senior Product Manager, HeadRush

This blog post is a high-level announcement. For a comprehensive breakdown of A2 visit NAM A2: The Complete Guide

The story behind A2

NAM launched as an open-source project in 2019 and quickly became the most accurate amp modeling technology available. The first iteration, Architecture 1 (A1), was designed by Steve Atkinson to run inside DAWs on computers, not embedded devices like multi-fx pedals.

A2 was developed by TONE3000 in partnership with Steve. Together, we rebuilt the architecture from the ground up to deliver better sound quality while using far fewer computational resources, opening the door for affordable multi-fx pedals and amps to run NAM natively on device.

A2 is the most accurate amp modeler ever

A2 captures sound virtually indistinguishable from the analog original. The bloom of a tube amp pushed into breakup, the sag of a fuzz pedal under a heavy chord, the snap of a transient through an analog compressor: A2 captures it all.

A2 represents dynamics, gain structure, frequency behavior, and transient handling more faithfully than any modeler that has come before it.

We verified this in two ways: head-to-head quantitative testing and large-scale blind listening tests.

Quantitative tests

We compared recordings of real gear against neural models from Neural Amp Modeler A2 and A1, Neural DSP Neural Capture V2, IK Multimedia's ToneX and Line6's Proxy. Each model was scored on how closely its output matched the original recording using four metrics: ESR, MAE, LOG_MEL and MRSTFT. Those scores were used to create head-to-head comparisons between models, using a single Bayesian Elo rating.

The Bayesian Elo score worked like a win/loss rating, where higher scores meant the model more consistently matched the real recording. The evaluation dataset spanned 39 tones, covering guitar and bass amps, amp and cab rigs, pedals, pedal and amp chains, outboard gear, and more. We modeled everything from sparkling Fender cleans to a vintage Neve 1073 to a dimed Mesa Boogie Dual Rectifier.

Blind listening tests

We also ran a large-scale blind listening test using the MUSHRA methodology, the audio industry's gold standard for evaluating perceived sound quality, often used by organizations like the BBC and EBU.

Over 1,000 participants submitted more than 100,000 ratings, which might be the largest tone study ever. In each trial a participant heard the real recorded gear first as a reference, then rated several subsequent versions of that same tone, modeled with A2-Full, A2-Lite, A1-Standard, Neural DSP’s Neural Capture V2, IK Multimedia’s ToneX, and Line 6’s Proxy, without knowing which was which. Participants scored each model on how closely it matched the original recording, using a 0–100 scale.

Both the test code and the raw ratings are open source:

• Test code: github.com/tone-3000/t3k-mushra
• Raw data: github.com/tone-3000/a2-mushra-data

A2 works anywhere

Until now, tones have been trapped in walled gardens. Audio plugins lock your tones into a plugin in your DAW. Neural DSP Quad Cortex, ToneX / Amplitube, and Line 6 Helix lock your tones into their pedals and plugins.

NAM files work anywhere NAM is supported, just like IR or MIDI files. This includes a rapidly growing ecosystem of plugins, pedals and amps. To accelerate this adoption, A2 can be run in two sizes:

• A2-Full is our maximum-accuracy model, built for pro audio. It delivers better quality than A1 and uses 30-40% less CPU. This means you can run three A2-Full models for the same CPU as two A1-Standard models.
• A2-Lite is designed for embedded devices like guitar and bass multi-fx pedals and amps. It runs at 50% CPU on a $3 ARM Cortex-M7 600MHz chip. A2-Lite makes TONE3000 possible on mass-market hardware.

To put that in perspective, a MacBook with an Apple M-series chip can run 64 A2-Full models or 200 A2-Lite models at once.

"Rather than build a walled garden where Blackstar tones only work on Blackstar gear, we chose to partner with TONE3000 so you can take your tones anywhere. This empowers you to take the sound in your head anywhere."
– Alex Gee, Head of Product at Blackstar

Fully open source

A2 is fully open source, creating an open protocol for tone that any company can build on. The model architecture, training code, and inference engine are free to use, modify, and ship in commercial products. Each new developer working with NAM joins a global community that's actively contributing to the ongoing development of A2.

TONE3000 API

Dozens of hardware and software companies are integrating with the TONE3000 API, allowing their users to browse and load tones from our library directly within their products. This includes access to hundreds of thousands of tones, including Neural Amp Modeler (NAM) captures, all available as the latest A2 models, plus our huge impulse response (IR) library. Every new tone shared on TONE3000 becomes instantly available across every connected product.

If you're a hardware or software maker interested in offering TONE3000 in your product, you can learn more here: TONE3000 API

From the team

"TONE3000’s mission is to make music creation universally accessible. A2 accelerates that mission by democratizing tone. A $5,000 vintage amp that was once locked away in a studio can now be captured, shared, and played by any artist, on any device, anywhere in the world."
– Stanley Vergilis, Co-Founder and CEO at TONE3000

"The goal was simple. Capture tone so faithfully that you couldn't tell the difference. Getting there took months of obsessing over every detail, from the odd harmonics of dimed EL34's to the sag of a vintage transformer. The entire universe of analog gear is now available to all."
– Woodbury Shortridge, Co-Founder and CTO at TONE3000

Thank You

Thank you for being part of this community. A2 exists because of the captures you've shared, the feedback you've given, and the years of work this community has put into Neural Amp Modeler and TONE3000.

A special thank you to the 1,000+ people who participated in our blind listening tests!

Learn More

We've published a comprehensive guide that contains everything you need to know about Architecture 2: NAM A2: The Complete Guide