CallCleaner distills SAM-Audio into a 19M bioacoustic denoiser
Original title: CallCleaner: Lightweight Bioacoustic Denoising from an Audio Foundation Model
Authors say the CPU model runs about 90 times faster than real time and leaves 11 dB less noise than Biodenoising.
The paper uses the source-separation model SAM-Audio, prompted with spans from a simple animal voice-activity detector and without training, to denoise field recordings of animal calls. They distill that approach into CallCleaner, a 19-million-parameter denoiser that needs no detector. The authors release code and weights for local denoising of single-caller recordings on a laptop CPU.
Unconfirmed: Speed and noise figures are the authors' measurements, not independently verified.
- CallCleaner has 19 million parameters and needs no detector at inference.
- Authors report about 90 times faster than real time on a laptop CPU.
- They report 11 dB less noise around the call than Biodenoising.
- They argue SI-SDR penalizes regenerated output and mixes noise removal with call fidelity.
From the original article
1 min read at arxiv.orgMany studies of animal vocalizations require clean recordings, but animals are usually recorded in the field, with background noise from weather, vehicles, and other animals. Animal vocalization denoising is challenged by the scarcity of clean training data and the extreme diversity of animal vocalizations.
By Michael Xie · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.