Biodenoising—A Novel Method for Noise Reduction In Animal Vocalizations
By Marius Miron, Sara Keen, Jen-Yu Liu, Benjamin Hoffman, Masato Hagiwara, Olivier Pietquin, Felix Effenberger, Maddie Cusimano
Ethologists and bioacousticians are constantly battling noise. Irrelevant signals – like the sound of wind or rain, interference from snapping shrimp, or even the animals’ own movements – can drown out the vocalizations they’re trying to analyze. Even in laboratory settings, background noise such as the whirrs and hums of fans and HVAC systems is a common challenge. This noise pollution can severely impact the accuracy of research, leading to incomplete datasets, misinterpretations, and missed opportunities for discovery.
In the human domain, denoising relies on large volumes of clean, isolated speech data. For animal vocalizations, however, this data simply doesn’t exist. The diversity of sounds from different animals and recording environments adds to the challenge, with existing models often struggling to generalize across different species or conditions. Add to that the additional challenge that some types of environmental noise provide important clues for understanding animal communication and behavior. For example, a herd of elephants charging or the sound of hoofbeats could provide important contextual clues, while the sound of moving water in a stream is less relevant.
Introducing Biodenoising: A Novel Solution
We have introduced a new method, Biodenoising, to address these challenges by denoising animal vocalizations without requiring clean data.
This is achieved by leveraging existing speech enhancement models, which have been trained on vast amounts of human speech data and possess a fundamental understanding of audio time series patterns. Instead of creating a new model from scratch, Biodenoising adapts these existing models to the specifics of animal vocalizations. This means that models can be trained on pseudo-clean targets created from pre-denoised vocalizations and segments of background noise.
Biodenosing was recently accepted to ICASSP and we’ve published a preprint on arXiv outlining our approach and introducing several key contributions:
- A new benchmarking dataset of clean animal vocalizations from diverse taxa, along with environmental noise samples.
- A separate training dataset comprising noisy vocalizations from existing bioacoustic datasets combined with environmental noise samples. This dataset is accessible via our Python library, biodenoising-datasets.
- A methodology leveraging pre-trained speech enhancement models to denoise animal vocalizations. Researchers can reproduce our experiments using the biodenoising Python library.
Biodenoising in Action
The performance of Biodenoising has been tested on a new benchmarking dataset, Biodenoising_validation, with 62 pairs of clean animal vocalizations and noise excerpts. Our tests demonstrate how Biodenoising maintains vocalizations while minimizing noise from fans, wind, and self-noise. Even in challenging underwater conditions, where signal-to-noise ratios are often very low, Biodenoising performs reasonably well. It's also being used to improve recordings that have previously been cleaned using speech enhancement models.
Our experiments also revealed that time-scaling animal vocalizations—slowing them down or speeding them up—significantly improved denoising performance. This approach takes advantage of the correlation between an animal’s body size and the pitch of its vocalizations. By shifting animal vocalizations into the human speech range, we can enable speech enhancement models to apply their learned signal priors, such as vowel and phoneme structures, to animal data.
Get Access
Researchers interested in exploring the method can access the pre-print, code, and libraries associated with Biodenoising. This tool is designed to open new doors for analyzing vocalizations in their natural contexts, revealing patterns that deepen our understanding of interspecies communication. Together, we can refine these methods and spur innovation at the intersection of AI and the natural world.
Join us in this journey. Access Biodenoising on GitHub and help shape the future of discovery, collaboration, and our relationship with the living world.
