@happysoftware/nam-web

Neural Amp Modeler as a drop-in AudioWorkletNode — no React, no COOP/COEP headers.

Run the real NAM — neural guitar amp models — anywhere in a Web Audio graph. It's just an audio node: connect() it, swap models at runtime, done. No UI framework to adopt, and — because inference runs inside the worklet, with no SharedArrayBuffer — no cross-origin-isolation headers to add to your site. Embed it on any page as-is.

npm install @happysoftware/nam-web

import { createNamNode } from '@happysoftware/nam-web';

const ctx = new AudioContext();
const nam = await createNamNode(ctx, { modelUrl: '/models/my-amp.nam' });

const guitar = ctx.createMediaStreamSource(stream);
guitar.connect(nam.node).connect(ctx.destination);

nam.setInputGainDb(6); // drive
await nam.loadModel('/models/other-amp.nam'); // hot-swap, glitch-free

Three lines to put a neural amp model in the signal path. No React, no UI, no SharedArrayBuffer headers — just an AudioWorkletNode.

The worklet and the WASM binary ship inside the package (dist/worklet.js, dist/nam-wasm.wasm). Serve them as static assets; the bytes are handed to the worklet at runtime — no special headers required.

Why

NAM in the browser already exists — TONE3000 ships it as a React component (neural-amp-modeler-wasm). But there is no headless, typed, framework-agnostic building block: an AudioWorklet node you can drop into a Vue/Svelte/vanilla-JS site, a web DAW, a model marketplace preview button, or a "hear this amp" product page.

nam-web is that missing layer: the engine only, three lines to wire up.

API

createNamNode(ctx, opts) returns a NamNode:

| Member | Purpose | |---|---| | node | the AudioWorkletNode — connect()/disconnect() as usual | | loadModel(url \| ArrayBuffer) | load/swap a model; resolves with its sample rate | | setInputGainDb(db) / setOutputGainDb(db) | drive (pre-model) and output level | | setBypass(on) | dry passthrough without tearing down the node | | on(fn) | subscribe to ready / loaded / error events | | dispose() | disconnect and release |

Lower-level pieces (NamModel, NamProcessorCore, RingBuffer, LinearResampler) are exported too, for offline rendering or custom graphs.

How it works

.nam model file ─► WASM (NeuralAmpModelerCore + Eigen, Emscripten)
                        ▲ lock-free ring buffer
AudioContext ──► AudioWorkletProcessor ──► output
                        ▲
            TypeScript API (createNamNode, loadModel, events)

• DSP core: upstream sdatkinson/NeuralAmpModelerCore, unmodified.
• WASM glue: this project's own Emscripten bindings (clean-room).
• TS layer: AudioWorklet wrapper, model loading, resampling, gain staging.
• No UI: the package ships zero components.

Source, demo, and build instructions: github.com/HappySoftwareLabs/nam-web.

Credits & licensing

• The neural inference core is NeuralAmpModelerCore by Steven Atkinson, MIT licensed. This project would not exist without it.
• The browser-WASM approach was pioneered by TONE3000 in their fork. nam-web is an independent, complementary implementation (headless engine vs. full player component).
• MIT licensed.