home/ modules/ deploying-neural-audio-into-max-and-pd

Deploying Neural Audio Models into Max/MSP and Pure Data

L2First instrument 4h K · AI & real-time generative AV 16 atoms

Learning objectives

learner can load a TorchScript model as an nn~ object and route audio through its methods
learner can export RAVE for nn~/VST and control it live via attributes and latent splitting
learner can reason about circular-buffer latency and waveform-to-waveform pass-through

Capstone — one whole task that evidences the objectives

Build a Max/MSP (or PD) patch that runs your exported RAVE model live: load the .ts via nn~, split encode/decode to manipulate individual latent dimensions in performance, expose model attributes as live controls, and document the circular-buffer size you chose, the latency it adds, and how the forward path behaves as a waveform-to-waveform effect. Also demonstrate the same model as a RAVE VST in a DAW.

Prerequisite modules

Training a RAVE Timbre Model End to End

This module crosses the gap that stops most people who train a RAVE model: getting it out of Python and onto a stage. The whole task is a performance-ready Max/MSP or Pure Data patch where your own timbre model runs live — the setting is an improvised electronic set where you sing or play into the model and sculpt its output in real time, with the same model doubling as a VST insert in a DAW session. TorchScript is the lingua franca here: understanding why a .ts file needs no Python is what makes the whole deployment story make sense.

The arc starts fully supported. First, install nn~ and load a pre-trained model (see “nn~ is a Max/PureData external that bridges trained neural audio models”) and get sound through the default forward path (“The forward method runs a neural model as an audio effect”) — a waveform-to-waveform pipe, audio in and audio out of the same length. Then swap in your own exported model, and open the hood: replace forward with separate encode/decode so the latent dimensions appear as signal outlets you can slider, LFO, and cross-fade (“Splitting RAVE encode and decode in nn~”). Add attribute messages for live control, then make a deliberate engineering choice — buffer size versus latency — which you must document in the capstone.

Every required atom gates that patch: methods, attributes, the circular buffer, waveform-to-waveform pass-through, and the VST workflow are each exercised directly. Supporting atoms widen the picture — the in_ratio/out_ratio contract that explains why latents run slower than audio, hot-swapping models mid-set with void mode, pre-trained model packs, offline batch generation, and ONNX as an alternative export — useful, but the capstone stands without them.