# Neural amp

> Play a bounded NAM A1/A2 WaveNet capture with shared DSP, verified model assets and explicit dry fallback when a model is missing.

Source: https://docs.tune.toys/nodes/std.nam/

Neural amp plays an attached `.nam` capture through the shared audio engine.

## What it does

The compiler admits WaveNet A1/A2 and containers of up to four WaveNets. Parsing and preparation happen before playback. Model files are limited to 8 MiB, finite weights and bounded topology. Unsupported architectures and features are refused with a `NAM_*` diagnostic. A missing model passes dry audio and reports `NAM_MODEL_UNAVAILABLE`.

## When to use it

Load a capture you have permission to use for an amp or pedal tone. Add Convolver after a capture that needs a cabinet impulse response. Local file uploads are private by default. TONE3000 catalog selection remains disabled pending the commercial agreement.

## Ports

<!-- ports:start -->

| Port    | Direction | Signal | Lanes                   |
| ------- | --------- | ------ | ----------------------- |
| `audio` | Input     | Audio  | Inferred from the cable |
| `audio` | Output    | Audio  | Inferred from the cable |

<!-- ports:end -->

## Parameters

<!-- parameters:start -->

| Parameter   | Label     | Range     | Default | Unit | Scale  | Notes                                                                           |
| ----------- | --------- | --------- | ------- | ---- | ------ | ------------------------------------------------------------------------------- |
| `input`     | Input     | -24 to 24 | 0       | dB   | linear | Input for the loaded neural amp model.                                          |
| `output`    | Output    | -36 to 12 | 0       | dB   | linear | Output for the loaded neural amp model.                                         |
| `normalize` | Normalize | 0 to 1    | 1       |      | linear | Normalize for the loaded neural amp model. Choices: 0 Off, 1 On.                |
| `size`      | Size      | 0 to 1    | 0       |      | linear | Size for the loaded neural amp model. Choices: 0 Full, 1 Lite.                  |
| `channels`  | Channels  | 0 to 1    | 0       |      | linear | Channels for the loaded neural amp model. Choices: 0 Mono (sum), 1 Per channel. |
| `mix`       | Mix       | 0 to 1    | 1       |      | linear | Mix for the loaded neural amp model.                                            |

<!-- parameters:end -->

<!-- data:start -->

- `model`: asset.

<!-- data:end -->

### Input (`input`)

Input trim changes how hard the model is driven, from −24 to +24 dB.

### Output (`output`)

Output trim adjusts the processed level, from −36 to +12 dB.

### Normalize (`normalize`)

On uses available model loudness metadata to target −18 dB. Without that metadata the gain is unity.

### Size (`size`)

Full chooses the largest submodel; Lite chooses the smallest. A new size warms for its receptive field and crossfades over 10 ms. A single model has the same tone for either choice.

### Channels (`channels`)

Mono averages input lanes and shares that result. Per channel gives each lane its own independent model state.

### Mix (`mix`)

Blend dry and processed audio. When sample rates differ, a prepared windowed-sinc resampler runs the model at its own rate; the dry path and graph compensation use the reported whole-sample latency.

## Tips

- Load a `.nam` file in the inspector or choose a model already attached to the Toy.
- Leave headroom at the output; captures differ in level.
- Preserve the creator's license and attribution when sharing.
- TONE3000 models cannot be bundled in releases or independent plugin exports.

## See also

[Drive](https://docs.tune.toys/nodes/std.drive/), [Convolver](https://docs.tune.toys/nodes/std.convolver/), [Gain](https://docs.tune.toys/nodes/std.gain/).