PoDAR: Power-Disentangled Audio Representation for Generative Modeling

TL;DR

PoDAR introduces a framework that disentangles signal power from semantic content in audio representations, enhancing modelability, accelerating convergence, and improving performance in generative audio tasks.

Contribution

It proposes a novel power disentanglement method using augmentation and consistency objectives, improving latent space modeling for audio generation.

Findings

Achieves 2x faster convergence with comparable quality.

Increases speaker similarity by 0.055 and UTMOS by 0.22.

Extends stable guidance regime to higher scales.

Abstract

The performance of audio latent diffusion models is primarily governed by generator expressivity and the modelability of the underlying latent space. While recent research has focused primarily on the former, as well as improving the reconstruction fidelity of audio codecs, we demonstrate that latent modelability can be significantly improved through explicit factor disentanglement. We present PoDAR (Power-Disentangled Audio Representation), a framework that utilizes a randomized power augmentation and latent consistency objective to decouple signal power from invariant semantic content. This factorization makes the latent space easier to model, which both accelerates the convergence of downstream generative models and improves final overall performance. When applied to a Stable Audio 1.0 VAE with an F5-TTS generator, PoDAR achieves about a $2\times$ acceleration in convergence to match baseline performance, while increasing final speaker similarity by 0.055 and UTMOS by 0.22 on the LibriSpeech-PC dataset. Furthermore, isolating power into dedicated channels enables the application of CFG exclusively to power-invariant content, effectively extending the stable guidance regime to higher scales.