shlee-97/rlat

Exploring the Design Space of Representation Learning for Audio Transformations

Paper · Code and documentation

Usage

Install the RLAT package, then encode precomputed latents shaped (batch, 64, frames):

import torch
from rlat import load_model

model = load_model("default", device="cuda")
with torch.inference_mode():
    embeddings = model(wet_latents.to("cuda"))
z_t, z_y = embeddings["z_t"], embeddings["z_y"]

Both embeddings have shape (batch, 1024).

For waveform input, use model.encode_audio(audio, input_sr=sample_rate), as shown in the quickstart. Inputs are resampled to 44.1 kHz when needed. Waveform encoding requires access to the Stable Audio Open audio codec; accept its access terms separately.