[KongchangAI]
· 1 min read· 996 words

CallCleaner: A New Breakthrough in Lightweight Bioacoustic Noise Reduction

CallCleaner: A New Breakthrough in Lightweight Bioacoustic Noise Reduction

Distilling SAM-Audio yields CallCleaner with 19M parameters, 90x real-time denoising on CPU.

Field animal recordings are often disturbed by background noise from weather, vehicles, etc. Clean training data is scarce and animal sounds are diverse, making denoising difficult. Traditional spectral gating leaves residual noise, and specialized neural bioacoustic denoisers have limited noise removal capability. A new study leverages the general-purpose audio separation foundation model SAM-Audio, which removes more noise without training and using only simple animal VAD to provide temporal hints, and distills it into CallCleaner with 19M parameters, running at approximately 90x real-time on a laptop CPU without a detector. Evaluation notes the limitations of the SI-SDR metric; using separate metrics, CallCleaner leaves 11 dB less noise than Biodenoising and dominates the Pareto frontier of noise removal versus call fidelity. Code and weights are open-sourced.

Title

CallCleaner: A New Breakthrough in Lightweight Bioacoustic Noise Reduction

Meta Description

CallCleaner is a lightweight bioacoustic denoiser with only 19M parameters, running 90x real-time on CPU, no detector needed, removing 11 dB less noise than Biodenoising, with open-source code and weights.

Article Content

Animal vocalizations in field recordings are often obscured by background noise from weather, vehicles, and other animals, and the difficulty of obtaining clean training data combined with the extreme diversity of animal sounds makes denoising a challenge. Traditional spectral gating methods leave residual noise, especially without voice activity detection (VAD); detector-free neural denoisers designed specifically for bioacoustics, while requiring no VAD, have limited noise removal capability. A new study shows that the general-purpose audio separation foundation model SAM-Audio, without any training and using only simple animal VAD to provide temporal hints, can remove more noise than either of the above two approaches. To make this heavyweight method practical, the researchers distilled it into CallCleaner—a denoiser with only 19 million parameters, no detector required, running at approximately 90x real-time on a laptop CPU.

CallCleaner Research Overview

From Heavyweight Model to Lightweight Tool: Distillation Strategy

SAM-Audio, as a general-purpose audio separation foundation model, possesses powerful source separation capabilities, but its computational overhead is large, making real-time local deployment difficult. The researchers employed knowledge distillation to transfer SAM-Audio's capabilities into a small student model, CallCleaner. With only 19 million parameters, this model achieves approximately 90x real-time processing speed on CPU and requires no detector whatsoever. This means users can quickly complete denoising on an ordinary laptop without a GPU or specialized hardware, while also ensuring privacy—all processing is performed locally.

Knowledge distillation is a model compression technique: a pre-trained large "teacher model" produces outputs (soft labels) on a large number of inputs, and a small "student model" simultaneously learns from both the true labels and the teacher's output distribution, thereby approximating the teacher's behavior with far fewer parameters. In audio separation tasks, the student model learns how the teacher estimates the target sound source from a mixed signal. Because the student model has a more compact architecture and does not need to run the teacher network at inference time, it significantly reduces computational and memory overhead, making real-time, local deployment possible.

Animal VAD (voice activity detection) in bioacoustics is used to determine when a target animal call is present in an audio segment, providing temporal segment hints. Unlike speech VAD, animal calls vary greatly in frequency band, duration, and pattern, and typically require simple detectors customized for a species or taxon. The approach used in this paper requires only such lightweight temporal hints to guide the separation model's attention to the periods when calls occur, without requiring precise frame-by-frame masks, thus reducing dependence on detection accuracy and maintaining the system's "detector-free" lightweight character.

Performance Comparison: The Trade-off Between Noise Removal and Call Fidelity

When evaluating denoising effectiveness, the traditional metric SI-SDR (scale-invariant signal-to-distortion ratio) has limitations: it penalizes regenerative outputs and conflates noise removal with call fidelity. The researchers adopted separate metrics to reveal the trade-off between the two. Experiments show that CallCleaner leaves 11 dB less noise around calls than Biodenoising, the leading neural bioacoustic denoiser with publicly available weights. Among fast, CPU-capable systems, CallCleaner dominates on the Pareto frontier of noise removal versus call fidelity, without requiring a detector.

SI-SDR (scale-invariant signal-to-distortion ratio) measures the similarity between an estimated signal and a reference signal after allowing for scale scaling; higher values indicate less distortion. However, it is based on point-by-point waveform comparison, is extremely sensitive to phase or time-shift changes introduced by generative models, and conflates "residual noise" and "target distortion" into a single score. Therefore, in denoising evaluation, SI-SDR may penalize regenerative outputs that sound good and cannot separately answer "how much noise was removed" and "how much of the call was preserved." Only by using two independent metrics—noise suppression and target fidelity—can the true position of a method on the Pareto frontier be characterized.

The Pareto frontier refers to the set of solutions in multi-objective optimization that cannot be improved in one objective without sacrificing another. In denoising, it corresponds to the optimal trade-off curve between "noise removal" and "call fidelity": methods on the curve have achieved the best current compromise, while methods below the curve have room for simultaneous improvement. Projecting different systems onto this plane allows intuitive comparison of which removes more noise at equal fidelity, or which achieves higher fidelity at equal residual noise.

Practical Significance and Open-Source Contributions

CallCleaner is designed specifically for single-vocalizer recordings, supporting fast, privacy-preserving local denoising without specialized hardware. The researchers have open-sourced the code and weights, making it easy for researchers and developers to use directly or further improve. This work demonstrates how the capabilities of general-purpose audio foundation models can be condensed into a practical tool, advancing the efficiency of field recording processing in bioacoustic research.

CallCleaner Schematic

As animal behavior research and biodiversity monitoring raise their demands for audio data quality, lightweight, detector-free denoising solutions like CallCleaner are expected to become a new addition to the toolbox of ecologists and acoustic researchers.

One-line Summary

Distilling SAM-Audio yields CallCleaner with 19M parameters, 90x real-time denoising on CPU.

Paragraph Summary

Field animal recordings are often disturbed by background noise from weather, vehicles, etc. Clean training data is scarce and animal sounds are diverse, making denoising difficult. Traditional spectral gating leaves residual noise, and specialized neural bioacoustic denoisers have limited noise removal capability. A new study leverages the general-purpose audio separation foundation model SAM-Audio, which removes more noise without training and using only simple animal VAD to provide temporal hints, and distills it into CallCleaner with 19M parameters, running at approximately 90x real-time on a laptop CPU without a detector. Evaluation notes the limitations of the SI-SDR metric; using separate metrics, CallCleaner leaves 11 dB less noise than Biodenoising and dominates the Pareto frontier of noise removal versus call fidelity. Code and weights are open-sourced.

#CallCleaner#bioacoustic denoising#SAM-Audio#knowledge distillation#lightweight denoiser#noise reduction
Share:

Related articles