Neural Models
Four LSTM-based amp models trained on real hardware. The most realistic distortion tones in 0xDrive.
How Neural Models Work
Each neural model is a Long Short-Term Memory (LSTM) recurrent neural network trained on paired dry/wet recordings captured directly from real amplifier hardware. The network learns the complete nonlinear, time-dependent transfer function of the physical device - including the subtle interactions between gain staging, output transformer saturation, and speaker impedance loading that classical DSP models approximate.
At runtime, inference is handled by RTNeural, a compile-time-optimised neural inference library designed for real-time audio. The LSTM forward pass runs entirely on the CPU within the oversampled processing block, with no memory allocation on the audio thread.
Training
Thousands of paired dry/wet audio samples captured from physical hardware at multiple drive levels and sample rates. LSTM weights optimised to minimise mean squared error across the full frequency spectrum.
Export
Trained weights are exported to RTNeural JSON format and embedded directly in the plugin binary as static data - no external model files, no disk access at runtime.
Inference
RTNeural runs compile-time-specialised LSTM layers with SIMD acceleration. The model processes one sample at a time within the oversampled loop, maintaining the stateful recurrent context between blocks.
The Four Models
Heavy, aggressive amp tone. The highest-gain model in the neural set - trained on a British-style high-gain amplifier head driven well into power-amp saturation. Produces thick, compressed distortion with pronounced low-mid weight and a scooped upper midrange characteristic of the source hardware.
Tight, focused mid-range drive. Trained on a Class-A boutique overdrive amplifier known for its exceptional note separation and pick attack clarity. Neural Spine emphasises the 800 Hz - 2 kHz presence band, delivering a forward, cutting tone that sits precisely in a dense mix without excessive low-end mass.
Searing lead tone with sustained harmonic content. Trained on an American-style combo amplifier with the gain pushed to the edge of feedback. Neural Venom captures the elevated upper harmonics, smooth compression, and singing sustain of a cranked lead amp - optimised for single-note lines and solos that need to float above the mix.
Warm, musical breakup with dynamic sensitivity. Trained on a vintage low-wattage all-tube combo pushed into natural power-amp compression. Neural Soul responds convincingly to input level - soft playing produces clean, harmonically rich tones while harder picking or higher Input Gain drives it into saturated breakup. The most touch-responsive neural model.
Usage Tips
Oversampling Quality
Neural models benefit significantly from oversampling because the LSTM operates on a richer, alias-free signal. Use 4x or higher for mixing sessions. At 1x, aliasing can interact with the model's harmonic output in unpredictable ways.
Recommendation: 4x for mixing, 8x for final render.
Driving the Model
The Input Gain and Drive knob work together to control how hard the LSTM receives signal. Input Gain scales before the model; Drive applies an internal multiplier (0.5× - 3.0×) to the model's input. For heavier saturation, raise Input Gain first, then increase Drive.
For clean tones: low Input Gain + low Drive (N.SOUL responds best here).
Parallel Blending
The Mix control enables parallel distortion - blending the neural model output with the clean dry signal. At 30-50% mix, the neural model adds harmonic weight and saturation while the dry signal preserves the original transients. Effective on bass, drums, and full-mix master bus processing.
Try N.SPNE at 40% mix on bass for harmonic enhancement without tone change.
CPU Usage
Neural models consume CPU comparable to the Classic DSP and GIMZWARE proprietary algorithms. RTNeural's compile-time-specialised layers and SIMD optimisation make the LSTM forward pass efficient enough for real-time use even at 16x oversampling on modern hardware.
All four models have identical architecture (LSTM) and therefore identical CPU cost - model selection does not affect performance.
Recommended Signal Chain for Neural Models
The Low Cut before the model prevents bass-frequency intermodulation. The High Cut after removes any residual harshness from the LSTM output. Reduce Output Level to compensate for the gain added by saturation.
Technical Details
- Architecture: LSTM recurrent neural network, compile-time fixed topology via RTNeural.
- Model storage: Weights embedded as binary data in the plugin binary - no external files required.
- Inference: Sample-accurate, stateful processing. The LSTM hidden state persists between audio blocks, preserving temporal continuity.
- Drive scaling: A continuous multiplier of 0.5× - 3.0× is applied to the model input, controlled by the Drive knob. This extends the perceptual dynamic range of the Drive parameter across the full physical knob travel.
- Thread safety: Model inference runs exclusively on the audio thread. No locks are used - the active model is selected via an atomic index with acquire/release ordering.
- Latency: Neural model inference adds no additional latency beyond the oversampling delay reported to the DAW.