The Patchbay_

RAVE vs DDSP

Both learn an instrument’s timbre and let you play through it, and they are the only two tools here that do. They disagree about how, and the difference shows up in what material each handles well.

Short answer. Live performance, messy or polyphonic source material → RAVE. Clean monophonic instruments and interpretable, tweakable control → DDSP. Note RAVE is non-commercial; DDSP is Apache-2.0.

Side by side

RAVE DDSP
ApproachVariational autoencoder over raw audioDifferentiable DSP inside a neural net
LicenceCC-BY-NC 4.0 — code includedApache-2.0
Commercial useNoYes
Real-timeYes — built for itNot the focus
DAW / patcherMax/MSP and Pure Data via nn~, plus a VST betaMostly notebooks
Control surfaceLatent dimensions you can manipulateExplicit pitch, loudness, harmonics
Best source materialAnything coherent, including polyphonicClean monophonic instruments
Training GPU8–32 GB by configModest

Where they actually differ

Interpretable parameters versus a learned latent

DDSP models audio with actual DSP components — oscillators, filters, harmonic amplitudes — that remain differentiable. The parameters therefore mean something: pitch is pitch, loudness is loudness, and you can reach in and change them predictably. RAVE compresses audio into an abstract latent whose dimensions have no assigned meaning. You can manipulate them, and the results are often musically interesting, but you are exploring rather than dialling in a known value.

RAVE is built to be played

RAVE exports a TorchScript file that loads into nn~ for Max/MSP and Pure Data, with a streaming mode using cached convolutions specifically so it works on a live audio stream. There is also a VST beta from Forum IRCAM. DDSP has no comparable performance path — it lives in Python and notebooks.

DDSP is fussier about its input

Because it models harmonics plus filtered noise, DDSP is at its best on clean monophonic material — solo violin, a single voice, a flute. Feed it a polyphonic mix and its assumptions break down. RAVE is far more tolerant: it will learn whatever corpus you give it, provided the corpus is stylistically coherent.

The licence gap is decisive for commercial work

RAVE is CC-BY-NC 4.0, and unusually that covers the implementation itself, not only released weights — so even a model you train on your own audio is produced with non-commercial tooling. DDSP is Apache-2.0 throughout. For anything you intend to sell, this is the row that matters.

Which should you choose?

Choose RAVE when…

Choose DDSP when…

RAVE detailsDDSP detailsAll ML & generative

← All comparisons