CallCleaner distills an audio foundation model into a 19M-parameter denoiser: field animal recordings run about 90 times faster than real time on a laptop CPU
Synopsis
Field recordings of animal calls carry wind, rain, vehicles and other animals, while clean training data are scarce. CallCleaner distills the general-purpose audio source-separation foundation model SAM-Audio into an 18.9M-parameter waveform-to-waveform denoiser that needs no detector or prompt and runs about 90 times faster than real time on a laptop CPU. On the 47 clips of at most 30 seconds in the Biodenoising test benchmark, it leaves 11 dB less noise around the call than Biodenoising and is better on 44 of 47 clips.
Interpretation
Denoising field animal calls can now be handled by a small model that needs no detector and no GPU, computing its output from the input waveform rather than regenerating audio. Earlier options were classical spectral gates, which leave residual noise especially without voice activity detection, or neural denoisers trained for bioacoustics, which remove less noise. On the 47 clips of at most 30 seconds in the Biodenoising test benchmark, CallCleaner leaves 11 dB less noise around the call than Biodenoising and is better on 44 of 47 clips; normalized for in-call energy it still leaves 11.2 dB less, better on 42 of 47.
The route is to use a general source-separation foundation model as a teacher, then compress it into a small denoiser that sees only the input audio. SAM-Audio is a flow-matching diffusion transformer that generates target and residual streams as audio-codec latents and accepts text prompts and time spans as optional conditioning; with text alone it often selects the wrong target. Prompted with time spans from a simple animal voice activity detector, SAM-Audio removes the most noise of any system; distilled from those pseudo-labels, CallCleaner runs about 90 times faster than real time on a laptop CPU, as fast as Biodenoising.
Evaluation also changes: instead of SI-SDR alone, noise around the call, noise under the call and distortion of the call itself are measured separately. SI-SDR compares waveforms sample by sample, penalizing any deviation from the reference and conflating noise removal with call fidelity; passing clean calls through SAM-Audio's codec alone, with no separation, yields a median SI-SDR of only 7.2 dB. On the Biodenoising test benchmark, CallCleaner and Biodenoising tie at 13.0 dB SI-SDR, yet CallCleaner leaves 11 dB less noise around the call, so a single SI-SDR cannot show that trade-off.
Perspective
The demonstrated setting is recordings of a single salient foreground animal whose calls are separated by background noise, evaluated on the 47 clips of at most 30 seconds in the Biodenoising test benchmark, mixed with real field noise at three SNRs and resampled to 16 kHz for scoring. The authors propose extending to dense targets and conditioning on which caller to keep, and improving the heuristic voice activity detector.
SAM-Audio's training data are not publicly listed, and 18 of the 47 clean calls are in FSD50K, so it cannot be ruled out that the teacher saw these clips; however, CallCleaner's pseudo-labels came from separate field recordings, and it still removes more noise than Biodenoising on 26 of the 29 clips not in FSD50K. The authors also note that intrusive metrics are limited by reference quality, so a reference containing background noise makes denoising count as missing and underestimates performance; extremely low- and high-frequency calls and impulsive noise remain challenging; and the evaluation reflects human hearing, while animals may perceive denoised calls differently.