ITO-Master: Inference Time Optimization for Mastering Style Transfer

Interactive demo of Inference Time Optimization (ITO) for Music Mastering Style Transfer. The mastering style transfer is performed by a differentiable audio processing model, and the predicted parameters are shown as the output. Perform mastering style transfer with an input source audio and a reference mastering style audio. On top of this result, you can perform ITO to optimize the reference embedding $z_{ref}$ to further gain control over the output mastering style.

ITO pipeline

Step 1: Mastering Style Transfer

Source Audio $x_{in}$

Reference Style Audio $x_{ref}$

all output samples are normalized to -12dB LUFS

Output Audio y'

Normalized Source Audio

Predicted Parameters

Step 2: Inference Time Optimization (ITO)

ITO Reference Style Audio $x'_{ref}$ (optional)

Number of Steps for additional optimization

1 100

Optimizer

Learning Rate

0.0001 0.1

Loss Function

AudioFeatureLoss CLAPFeatureLoss

AudioFeatureLoss Weights (comma-separated)

RMS, Crest Factor, Stereo Width, Stereo Imbalance, Bark Spectrum

all output samples are normalized to -12dB LUFS

ITO Output Audio

ITO Step

1 100

ITO Predicted Parameters

ITO Log