FdAudio achieves one-step text-to-audio with 28.8% quality boost

A team of researchers from National Taiwan University (Kuan-Po Huang, Bo-Ru Lu, Ho-Lam Chung, Shih-Hsin Wang, Hung-yi Lee) has introduced FdAudio, a novel post-training framework for one-step text-to-audio generation. Existing few-step models like MeanAudio accelerate synthesis by modeling average velocities, but their strict one-step output quality lags behind multi-step counterparts. FdAudio closes this gap by directly optimizing the one-step distribution using a multi-representation Fréchet-distance (FD) loss across pre-trained embedding spaces, rather than relying solely on regression against target velocity fields. Critically, naive FD post-training causes multi-step degradation; FdAudio solves this with a MeanFlow consistency objective that acts as a structural anchor, ensuring high-fidelity audio even when sampling with 25 steps.

Experimental results show FdAudio establishes state-of-the-art one-step quality among few-step systems, achieving an 11.4% reduction in FD score and a 28.8% improvement in FAD score relative to the baseline MeanAudio framework. The MeanFlow anchor also enables 25-step sampling paths that match or surpass strong multi-step models at a fraction of their computational cost. This breakthrough makes real-time, high-quality text-to-audio generation feasible for applications like content creation, gaming, and accessibility tools, where low latency is critical.