PoDAR: Power-Disentangled Audio Representation for Generative Modeling

2026-05-11

AI Summary

This work addresses the challenge in audio latent diffusion models where signal power and semantic content are tightly coupled in the latent space, complicating effective modeling. To resolve this, the authors propose the first explicit decoupling of these two factors by introducing stochastic power augmentation and a latent consistency objective, thereby constructing a structurally cleaner and more tractable latent representation. This approach not only enhances training efficiency and generation quality but also enables classifier-free guidance (CFG) to be applied solely to semantic content, significantly improving stability under high guidance scales. Evaluated on the LibriSpeech-PC dataset, the method achieves approximately 2× faster convergence compared to the baseline, along with a 0.055 improvement in speaker similarity and a 0.22 gain in UTMOS score.

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.

Problem

• audio representation
• latent space
• factor disentanglement
• generative modeling
• signal power

Innovation

• power disentanglement
• latent diffusion models
• audio representation
• randomized power augmentation
• classifier-free guidance