New AI Cleans Up Noisy Audio Using 87% Less Computing Power
Your next phone call could be crystal clear — using a fraction of the energy.
A new approach called SlimDiffuSE makes diffusion-based speech enhancement far more efficient. While these models deliver state-of-the-art performance, they typically require many evaluations of a large neural network — driving up computational costs. SlimDiffuSE tackles this by using adaptive network widths throughout the data generation process, guided by a greedy search algorithm that optimizes the width schedule. The result: performance comparable to baseline diffusion models with up to 87.5% less computational complexity, and no significant drop in key objective metrics like PESQ and SI-SDR.
- Speech enhancement AI removes background noise from audio, but normally needs huge computing power.
- SlimDiffuSE shrinks or grows the AI's "brain" at each step, slashing computing costs by up to 87.5%.
- Sound quality stays almost the same as full-size models, making it practical for phones and hearing aids.
Why It Matters
Cheaper AI audio cleanup means clearer calls and better hearing devices, powered on small, everyday gadgets.