Loading [a11y]/accessibility-menu.js
Drift-DiffuSE: Diffusion Model with Learnable Drift Term for Speech Enhancement | IEEE Conference Publication | IEEE Xplore

Drift-DiffuSE: Diffusion Model with Learnable Drift Term for Speech Enhancement


Abstract:

Speech enhancement (SE) has become a crucial component of the front end of speech systems. In recent years, diffusion-based speech enhancement has made significant progre...Show More

Abstract:

Speech enhancement (SE) has become a crucial component of the front end of speech systems. In recent years, diffusion-based speech enhancement has made significant progress. However, existing diffusion-based speech enhancement methods primarily focus on diffusion terms in the diffusion model, overlooking drift terms. In this paper, we improve the Stochastic Differential Equation (SDE) and propose a novel framework for speech enhancement. First, we improve the modeling capabilities of our model by using two independent modules to fit the drift term and diffusion term in our SDE. Then, to explore the best way to combine modules, we have devised two distinct combination strategies. Finally, we also propose a novel loss function. The experimental results demonstrate the superior performance of our model in speech enhancement tasks, outperforming existing generative models in Perceptual Evaluation of Speech Quality (PESQ). Moreover, our approach achieves good speech enhancement performance even with fewer diffusion steps and exhibits good generalization capabilities, which can improve the efficiency of diffusion model-based applications. The code are available at https://github.com/finestu/Drift_DiffuSE.
Date of Conference: 30 June 2024 - 05 July 2024
Date Added to IEEE Xplore: 09 September 2024
ISBN Information:

ISSN Information:

Conference Location: Yokohama, Japan

Contact IEEE to Subscribe

References

References is not available for this document.