1024 text colours
1024 of them — 128 families × 8 weights: 1 is the strongest, 8 the softest. Each one is solved for the background it is used on, so it stays legible on white and on black. The list is in four sections: Neutral, Spectrum, Named and Flat.
1024 colour schemes
1024 of them — 128 families × 8 depths: depths 1–4 are light pages, 5–8 dark ones, and the family called Plain is white, greys and black. The Spectrum families are thirty-three evenly spaced stops around the hue circle, the Named ones are the everyday colours between them, and the Flat ones — tinta unita — are those same colours kept at full strength, mixed with nothing but white or black. A scheme’s colour is the same hue at every depth, only lighter or darker.
1024 typefaces
1024 typefaces — as many as the colour schemes, and chosen the same way: a long list rather than a short one, ranked by how much the web actually uses them. Every one is self-hosted, latin only, one file each, and open source. They are sorted into five sections — Sans Serif, Serif, Display, Handwriting and Monospace — and each row draws itself in its own face only once it is on screen, so opening this panel costs nothing and a typeface can never slow a page down.
Open-source audio · NAM · TONE3000
Open-source Neural Amp Modeler technology, combined with the free TONE3000 library of more than 700,000 captures and impulse responses, has changed how guitarists get their tones. As the creators of Chain of Sound, we build professional-grade signal chains entirely from the free TONE3000 plugin and community-shared NAM captures — no paid hardware and no paid software required.
Platform-published figures, repeated as claims rather than measurements. What we can verify ourselves — the format, the licence and the software in your package manager — is documented below, with the method.
01 · The foundation
Neural Amp Modeler is a fully open-source modelling technology created by Steven Atkinson. It trains a neural network on the way a specific piece of analogue gear — an amplifier, a pedal, a channel strip — responds to a signal, then runs that trained network as a plugin. The result is a measurement of one device rather than an approximation assembled from a schematic, a tone stack and a filter curve.
Training code, file format and player are readable, forkable and embeddable. No licence server, no per-seat cost, no dongle, no vendor deciding when your tone expires. That is why this ecosystem grew faster than any single company could have funded.
A capture is compared against the device it models with a test signal and an error metric. You can re-run the comparison on your own bench, which is the difference between a specification and an advertisement. “Sounds better” never enters the argument.
A .nam file runs on Linux, inside a DAW, on a floor unit or on a
single-board computer. Progress compounds across products instead of resetting with each
new generation of hardware.
On the blind test. The NAM community reports that listeners could not reliably tell a capture from the amplifier it models in blind A/B tests. We repeat that as a claim from the project’s own community, not as a peer-reviewed result. The defensible version is narrower and still useful: a good capture of a good amplifier is close enough that the model stops being the thing you notice in a mix. Test it on your own rig before you repeat it as fact.
02 · The library
A model format is only as valuable as the library behind it. TONE3000 is where that library lives: a free, community-uploaded catalogue of NAM captures and cabinet impulse responses. Since the launch of NAM A2 in June 2026 the catalogue has more than doubled — from roughly 300,000 to over 700,000 tones in about three months — alongside a reported 600,000 monthly active users and 9 million plugin downloads.
Clean platforms, edge-of-breakup, high-gain stacks and bass heads — usually captured per channel and per setting, which is what you want: a capture is a snapshot, so a useful library gives you several snapshots of the same amplifier.
Overdrives, distortions, fuzzes, boosts and the occasional compressor or preamp. These are the blocks you place before the amp model, and they matter more than most people expect when the goal is a chain rather than a single tone.
Complete rigs captured end to end, plus the cabinet and microphone impulse responses that turn a modelled head into something that sits in a mix. A cabinet IR is a linear system, which is exactly what convolution is good at.
Licensing
Etiquette
On ownership. Publishing a capture of an amplifier you own is publishing a measurement of a device, not the maker’s schematic or design. Credit the manufacturer, date the capture and document the signal path. Several manufacturers now do this themselves — see the hardware section.
03 · The tool
Launched on 1 September 2026 and free of charge, the plugin is both the front end to the catalogue and the place where the chain gets built. You search the library, audition captures, stack them, match levels and print a result without leaving the DAW or opening a browser.
| Feature | What it means in practice |
|---|---|
| Built-in tone browsing | Search the whole catalogue and load a capture without leaving the session, which removes the download-folder chaos that used to surround NAM. |
| Unlimited signal chains | As many captures in series or parallel as the CPU allows. A pedalboard is a graph, not a fixed list of slots. |
| Dual chains & branching | Split the signal into two paths and rejoin them: clean and dirty, DI and amped, or two cabinets on one head — the normal way to work on bass. |
| Level-matched A/B | Global gain staging, EQ and normalisation between blocks, so the louder chain does not automatically win the comparison. |
| Full platform support | VST3, AU, AAX, CLAP, LV2 and a standalone build on macOS, Windows and Linux. The LV2 build is what makes this workflow possible on a stock distribution. |
Method
Start from the amplifier, because it sets the attack, the compression and the way the chain breathes. A pedal in front of the wrong amp model is a pedal in front of the wrong amp.
Drives and boosts before the amp, modulation and time-based effects after the cabinet. Keep every stage level-matched as you add it, or you will be choosing between chains by volume instead of by tone.
The cabinet and microphone position decide most of what people call “the tone”. Change the IR first, then reach for the EQ — in that order, so you are correcting the right thing.
Normalise, print, and write the recipe down: captures, versions, IR, gain settings and the input level. A chain you cannot rebuild next year is a session you will repeat from memory.
Recipe
The reference chain we build from most often. Every block comes from the free library; the only decision that costs money is the interface you already own.
| Block | Typical choice | Why it sits there |
|---|---|---|
| Front | Boost or overdrive capture | Tightens the low end before the amp sees it — the oldest trick in the book. |
| Amp | NAM capture of the head | Sets the character; only as accurate as the level you feed it. |
| Cabinet | 4×12 impulse response | Most of the frequency response of the finished sound lives here. |
| Space | Plate or room, low mix | Depth without smearing the transient you just spent an hour on. |
Calibrate the input before you judge a capture. This is the single most common reason a model sounds wrong. A capture was trained at a particular input level; feed it 6 dB hot and it distorts in a way the real amplifier never would, then everyone blames the model. Set the interface so a hard DI peaks around −12 to −6 dBFS, keep the plugin at unity, add no gain ahead of the amp block, and re-check the level whenever you switch captures. Fizzy, thin or blown-out is a level problem far more often than a modelling problem.
04 · Cabinets
NAM models the amplifier, which is a non-linear device. The cabinet, the microphone and the room are a linear system, and a linear system belongs in a convolution impulse response. That is why the same capture through two different IRs sounds like two different rigs — and why a chain is not finished until the cabinet is chosen deliberately.
Same microphone position, same speaker, same temperature where possible. Note the geometry; a metadata file turns a stray WAV into a reproducible measurement.
Cover the full band you care about, commonly 20 Hz to 20 kHz, long enough to excite the longest decay — typically 5 to 20 seconds.
Capture the sweep plus the full decay tail. Leave generous silence after the sweep so the reverb or the cabinet resonance is not truncated.
The Farina method: convolve the recording with the time-reversed, amplitude-compensated sweep. Harmonic distortion products arrive before the linear response and can be windowed out — which is the part a plain IR loader cannot do for you.
Trim to the direct sound, normalise consistently across the whole library, and store sample rate, microphone, gain, position and date alongside the file.
Tooling
-2.jpg){kind=link}
# 1. Generate a 10 s exponential sweep, 20 Hz - 20 kHz
sox -n sweep.wav synth 10 sine 20 \
synth 10 sine 20000 \
fade 0.01 10 0.01
# 2. Record the space (adjust device and channels for your rig)
sox -t alsa hw:1,0 -c 1 -r 48000 -b 24 capture.wav trim 0 12
# 3. Deconvolve by convolving with the inverse sweep
# (reverse it, then apply the amplitude envelope)
sox sweep.wav inverse-raw.wav reverse
sox inverse-raw.wav inverse.wav vol -6dB
sox capture.wav inverse.wav ir.wav
# 4. Inspect the result before you trust it
sox ir.wav -n stat
sox ir.wav -n spectrogram -o ir.pngThe exact inverse-filter envelope depends on your sweep parameters and playback chain. Verify by measuring a known reference — a straight cable loopback should produce an impulse response that is close to a single spike.
Encode what you cannot hear in a filename: maker, model, microphone, position, distance, date. 1960A-sm57-cap-2in-2026-03.wav tells you more than cab_final_v7.wav ever will.
Consistent nominal level across every file, no clipping in the capture, correct sample rate and bit depth, and one honest A/B against the real cabinet. Document the noise floor.
A clean measurement of a microphone position nobody would use is still a useless file. Publish the measurement and the opinion side by side, labelled as such.
05 · Hardware
A format becomes infrastructure when the people who build the analogue gear it models start shipping it. Each of these makers did something more useful than adding a logo to a compatibility list: they published captures, integrated the library, or built hardware around a catalogue they do not own.
The British amplifier builder publishes signed-off NAM captures of its own amplifiers to TONE3000, free of charge. That settles the “is modelling legitimate” argument with the strongest possible answer — the manufacturer’s own measurement — and it hands you a reference file: compare your capture, your IR and your gain staging against the maker’s version instead of against your memory of a room.
Prime, Core and Flex Prime integrate the TONE3000 API directly: browse and load community captures from the touchscreen with no computer anywhere in the signal path. For live work that changes the failure mode — the thing that can break at a gig is a floor unit, not a laptop — and the tone you auditioned at home is the tone you carry to the stage.
The Anagram runs multiple NAM A2-Full captures simultaneously with a built-in TONE3000 browser. Running several models in parallel is what makes this interesting for bass: keeping a clean low end under a driven top end is the standard technique, not an exotic effect, and it needs a device that can afford the CPU.
Case study
SonuLab is the clearest example of a maker treating NAM as a first-class format rather than a compatibility checkbox. Its floor unit is not a closed ecosystem with a modelling engine bolted on; it is a player for a catalogue that belongs to the community, and the engineering decisions follow from that.
The StompStation PRO reads NAM A2-Full files natively, which turns 700,000 community captures into a hardware instrument instead of a plugin that needs a computer kept alive beside your pedalboard. Up to two NAM files can run in a single preset, and the same patch carries a complete stereo chain — impulse-response loader, modulation, delay and reverb — so the pedals you would otherwise bring are already in the signal path. For a chain builder that is the difference between modelling a rig and performing with one.
SonuLab’s VXAMP engine runs its internal processing at an effective rate above 700 kHz inside a 192 kHz, 32-bit signal path, with a claimed 117 dB of dynamic range. On paper that is overkill; in practice it is how a digital rig avoids the aliasing and the thin “digital sheen” that make older modellers obvious on clean and edge-of-breakup material — which is precisely where captures live or die. Builders should read those numbers as an argument about artefacts, not as a marketing score.
VoidX-Control turns a phone or tablet into the editor: build the chain, swap captures, rename presets, keep the storage tidy and update firmware without balancing a laptop on a flight case. It is the part most hardware makers skip, and it decides whether you actually use a large library or stay on the first preset you liked. Editing is where a free catalogue becomes a usable one.
For the Chain of Sound workflow, SonuLab matters because it closes the loop: the same captures and the same chain you built on a desktop, running on hardware at the end of a cable. We are not affiliated with SonuLab and receive nothing for writing this.
| Device | Library access | NAM capability | Why it matters |
|---|---|---|---|
| Blackstar amplifiers | TONE3000, free | Maker-published captures | A manufacturer-verified reference to measure your own chain against. |
| HeadRush Prime / Core / Flex Prime | TONE3000 API, on-device | Capture player with a floor-friendly workflow | No computer in the signal path; the stage rig and the studio rig agree. |
| Darkglass Anagram | TONE3000 browser, on-device | Multiple A2-Full captures in parallel | Parallel bass chains without a laptop and an interface. |
| SonuLab StompStation PRO | Community catalogue, on-device | Two NAM files per preset, full stereo chain | An open catalogue as the instrument, with the pedals included. |
06 · Your distribution
“Free” here means two things. The captures are free, and so is most of the software you need to record, mix, model, master and measure them. A stock Linux distribution carries a complete audio stack in its package manager: a professional DAW, a modular patchbay, hundreds of LV2 effects, synthesisers, drum machines, a DJ application, a spectrogram tool and a notation engraver. You do not have to build a studio from source to build a signal chain.
Package names below were checked against the Ubuntu 24.04 repositories. Names differ between distributions and change between releases; a few tools ship only as an upstream build, a Flatpak or an AppImage, and those are marked as such.
| Layer | Package | What it gives you |
|---|---|---|
| Audio server & routing | pipewire, wireplumber | A low-latency graph with per-application routing that also speaks JACK, which is why modern releases do not need a separate JACK daemon to run pro audio. |
| Patchbay & plugin host | qpwgraph, carla | Drag-and-drop connections between applications and ports; Carla also hosts LV2, VST3 and CLAP plugins outside a DAW. |
| Multitrack DAW | ardour | Recording, non-destructive editing, automation and mixing — the reference open-source DAW, and where a Chain of Sound session normally lives. |
| MIDI-first sequencing | qtractor, rosegarden, lmms | Arrangement and MIDI work; LMMS for beat-driven production, qtractor when you want an LV2 host with a timeline. |
| Two-track editing | audacity | Fast destructive edits, format conversion, batch macro processing and quick spectrogram checks. |
| Analysis & metering | sonic-visualiser, japa, easyeffects | Spectrograms, loudness and spectrum analysis, plus system-wide EQ and limiting when you need to tame a room or a laptop speaker. |
| DJ & library | mixxx | Library management, timecode and controller support, and a second use for the same audio interface. |
| Job | Package | What it gives you |
|---|---|---|
| Amp & pedal modelling | guitarix, neural-amp-modeler (pip), AIDA-X (upstream LV2) | Guitarix as a ready-made guitar chain and NAM host; the NAM trainer to build your own captures; AIDA-X to run them as a plugin in any LV2 host. |
| Effects racks | calf-plugins, lsp-plugins, zam-plugins, x42-plugins, mda-lv2 | Compressors, EQs, reverbs, limiters, gates and meters — hundreds of LV2 plugins, all scriptable and all free of licence servers. |
| Synthesis & sampling | zynaddsubfx, yoshimi, samplv1, drumkv1, Surge XT (upstream) | Additive, subtractive and wavetable synthesis plus sample and drum instruments, so the chain is not only guitar. |
| Drums | hydrogen, drumgizmo | Pattern programming and multisampled acoustic kits for writing and rehearsing against a part that keeps time. |
| Live looping | sooperlooper, giada | Loop stacking for performance and practice; both run inside the same PipeWire graph as the rest of the session. |
| Capture training | sox, ffmpeg, neural-amp-modeler | Re-amping, file surgery and model training stay scriptable, which means a capture can be reproduced from a shell script instead of a memory. |
| Notation & scores | musescore3, lilypond | Engraving, playback and text-based scores that live in version control beside the session files. |
The shortest path to a working studio: a low-latency kernel, PipeWire and JACK configured to cooperate, and the audio meta-packages that install the toolchain in one transaction. If the goal is making sound rather than learning packaging, start here.
Fedora’s Jam lab, the Arch pro-audio group and openSUSE’s audio
patterns cover the same ground on a faster release cadence. They assume you read the
wiki and enjoy it — and they are where new PipeWire features land first.
A Raspberry Pi with a class-compliant interface runs a NAM host on a stock ARM image, which is where the audio track meets the open-hardware track. Be honest about the limit: it is a fine practice and recording rig, not yet a low-latency stage rig.
Install pipewire and wireplumber, keep the pipewire-jack
compatibility layer, and stop starting jackd by hand. Verify with
pw-top and wpctl status rather than by guessing.
Add your user to the audio group and let rtkit grant priority.
Raise memlock limits in /etc/security/limits.d/; then leave the
rest of the folklore alone until a measurement asks for it.
Watch for xruns at your working buffer size, not at the smallest size you can force. A slightly larger quantum you never notice beats a dropout you always do.
Training your own captures on Linux. The trainer installs from
pip and runs on a CPU, though a discrete GPU makes the wait tolerable — CUDA
and ROCm are both supported upstream. Re-amp through the interface, record the DI and the
amp in two takes of exactly the same performance, and keep the input level written down.
That level is the number the whole chain depends on.
ardour row in the first table.
Recorder, editor and mixer, three windows, no licence server and no account to keep alive.
A Chain of Sound session normally lives here, because nothing in it stops opening when a
subscription ends. The current release is Ardour 9.2; this screenshot is 6.7, and it is
the newest one anybody has released under a free licence — the alternatives are
from 2008, 2009 and 2013, and two of those are under licences we cannot reuse. An old
version we may show you beats a current one we may not.
Screenshot: Derwok —
CC0,
resized,
source.
audacity row in the first table.
The two-track editor that is already installed on most of the machines reading this page.
Trim a take, convert a format, run a macro across a folder of captures, look at a
spectrogram before you trust your ears. This screenshot is new rather than historical:
it was published under a free licence on 3 September 2026, which the Ardour note above
shows is not the normal case.
Screenshot: Zurel —
CC BY-SA 4.0,
resized,
source.
lmms end of the MIDI-first sequencing row.
Step sequencer, bundled synths, a demo song already loaded: the quickest way to find out
whether an idea is worth keeping before you open anything heavier. Released CC0, which
asks for nothing at all — we credit it anyway, because a credit is cheaper than an
argument.
Screenshot: Ranged Ranger —
CC0,
resized,
source.
07 · The guide
This page is the spine of the series. Each part below becomes its own document, written with the same rule as everything else we publish: state the method, show the measurement, and print the corrections where people can find them.
Install the plugin, calibrate the input, load one capture and A/B it against a clean DI. The first hour decides whether everything after it works.
A clean pedal platform, a high-gain stack, a parallel bass rig and an ambient chain — each built end to end from captures already in the free library.
Shortlists organised by job rather than by brand: edge-of-breakup, fuzz, bass DI, direct-to-desk. Judged on how they sit in a mix, not on how they sound alone.
What each floor unit really does with NAM — how many captures, which format version, how it browses the library — so you buy hardware for the workflow you actually have.
The people publishing captures: how they run a capture session, how they document it, and how they handle consent when the amplifier belongs to a client.
Our measurements, the captures that failed, and the impulse response that measured perfectly and sounded wrong. Published with the method, corrected in public.
08 · Questions
The library and the plugin are free to use, and community uploads are published under the platform’s licence so they can be shared at all. Curated commercial packs exist alongside them and buy you convenience and a smaller shortlist, not legality.
A capture is cheaper to run than a full physical-modelling suite, so an older laptop handles several blocks comfortably. A2-Full models and dual-chain patches cost more, so print the chain when a mix gets dense instead of buying a machine to audition it.
Yes, and that is the point of the distribution section. The plugin ships an LV2 build, captures are plain files, and the rest of the session — Ardour, PipeWire, the LSP and Calf plugins — comes from the repositories you already have.
No. It is a measurement of one amplifier, at one setting, through one signal path. Change the input level, the cabinet or the room and you have changed more than the model did. Treat a capture as a documented starting point, not as a copy of an object.
09 · Provenance
!.jpg){kind=link}
errata@polymathes.cc and it comes down.errata@polymathes.cc. They are reviewed in public, and the page is edited rather than quietly reposted.Contact
A capture we should hear, an IR that measured well and sounded wrong, a distro package that has moved, or a correction to something on this page — all of it is welcome. Measurements and session files beat opinions, but opinions are read too.
audio@polymathes.cc
Weekly on the mailing list
errata@polymathes.cc
Reviewed in public
Click anywhere, or press Esc, to close.