New AI model cleans noisy bioacoustic recordings with ridge-guided U-Net

Bioacoustic recordings are often buried in ambient noise, making it hard to analyze weak or overlapping vocalizations. While convolutional U-Nets have excelled at speech and music denoising, their direct use on bioacoustic data is stymied by a lack of clean training sets. To solve this, Reyhaneh Abbasi and colleagues proposed a training set synthesis approach that leverages frequency ridges—contours representing the fundamental frequency and harmonic partials of vocalizations. Their supervised model predicts a complex ratio mask in the time-frequency domain, and a novel ridge-guided loss function assigns higher weights to ridge regions, forcing the network to preserve vocalization detail during denoising. In a case study on house mouse ultrasonic vocalizations (USVs), the method outperformed a previous signal-processing approach in tracking fundamental and harmonic partials on actual field recordings.

Beyond raw denoising, the team showed that a classifier trained on denoised data improved USV classification on out-of-sample, noisy recordings from wild and domesticated mice compared with classifiers trained on noisy audio. On synthetic test data, the method substantially improved the scale-invariant signal-to-distortion ratio across a wide input signal-to-noise ratio range. Published in IEEE Transactions on Audio, Speech and Language Processing (vol. 34, 2026), the work explicitly targets behavioral biology and neuroscience applications, where USV analysis is critical. The authors argue the approach is broadly applicable to any bioacoustic signal with trackable ridges, opening the door to robust denoising tools for ecoacoustics and animal communication research.