Clearing Up Background Noise in Speech

November 2, 2026

Cleaning up audio to make speech clearer is a common challenge. Usually, AI models need pairs of clean and noisy speech to learn. Since these pairs are often made artificially, they do not always match real-world recordings. A method called noisy target training tries to fix this by training models on real, noisy recordings. However, this method can still leave some unwanted background noise in the final audio.

To solve this, researchers introduced a new technique called drifting loss. This method uses separate clean speech to guide the model. It pulls the enhanced speech closer to clean speech patterns while pushing it away from noisy ones. This helps remove the remaining background noise without needing perfect pairs of clean and noisy recordings.

Background noise is one of many things the quality of a dub depends on, along with how fast and clearly people speak and how many people speak at once. So some lines often need a fix, and for that you need a tool that is flexible and easy to use. On the Velotras Video adjustment page, which works even on a phone, you fix the dub line by line.

What you can correct in a dub, and how: Dubbing you can correct

Want your video dubbed into another language? Give it a try, for free: velotrasvideo.com

Source: Noisy Target Training Meets Drifting for Weakly-Supervised Speech Enhancement — Yoshiki Masuyama, Dominik Klement, Christoph Boeddeker, Julius Richter, Kohei Saijo, Gordon Wichern et al., arXiv, 2026