MeanFlowSE: one-step generative speech enhancement via conditional mean flow

2025-09-18

AI Summary

Real-time speech enhancement suffers from high computational overhead due to iterative ODE solvers in flow- or diffusion-based generative models. Method: This paper proposes MeanFlowSE, a single-step generative speech enhancement method built upon conditional mean flows. It uniquely models finite-time displacement as the average velocity field over trajectory intervals and derives a local training objective satisfying instantaneous field constraints—enabling one-step reverse generation without knowledge distillation or teacher models. Exact implementation of the MeanFlow identity is achieved via Jacobian-vector products (JVPs), with optional few-step refinement. Results: On VoiceBank-DEMAND, the single-step MeanFlowSE matches or surpasses multi-step baselines in intelligibility, fidelity, and perceptual quality (e.g., PESQ, STOI, DNSMOS), while accelerating inference by orders of magnitude and substantially reducing computational cost.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Security and Privacy: Privacy-enhancing technologies

Abstract

Multistep inference is a bottleneck for real-time generative speech enhancement because flow- and diffusion-based systems learn an instantaneous velocity field and therefore rely on iterative ordinary differential equation (ODE) solvers. We introduce MeanFlowSE, a conditional generative model that learns the average velocity over finite intervals along a trajectory. Using a Jacobian-vector product (JVP) to instantiate the MeanFlow identity, we derive a local training objective that directly supervises finite-interval displacement while remaining consistent with the instantaneous-field constraint on the diagonal. At inference, MeanFlowSE performs single-step generation via a backward-in-time displacement, removing the need for multistep solvers; an optional few-step variant offers additional refinement. On VoiceBank-DEMAND, the single-step model achieves strong intelligibility, fidelity, and perceptual quality with substantially lower computational cost than multistep baselines. The method requires no knowledge distillation or external teachers, providing an efficient, high-fidelity framework for real-time generative speech enhancement.

Problem

• Single-step generative speech enhancement for real-time applications
• Eliminates need for iterative ODE solvers in flow-based systems
• Maintains high intelligibility and fidelity with lower computational cost

Innovation

• One-step generative speech enhancement via conditional mean flow
• Uses Jacobian-vector product for finite-interval displacement supervision
• Single-step backward-in-time displacement without multistep solvers