🧬 biology

Acoustic Data Synthesis for Evaluating Noise Reduction in Bioacoustic Event Detection

This paper introduces an open toolbox for synthesizing marine acoustic recordings at controlled signal-to-noise ratios to evaluate noise reduction front-ends, demonstrating how such processing affects event detectability and the performance of pre-trained detection models.

Original authors: Bram Cuyx, Clea Parcerisas, Elisabeth Debusschere, Thomas Dietzen

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bram Cuyx, Clea Parcerisas, Elisabeth Debusschere, Thomas Dietzen

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the ocean as a giant, underwater concert hall. It's never silent; it's filled with a constant, churning roar called the "soundscape." This isn't just background noise; it's a mix of nature's own music (whales singing, fish clicking), the earth's rumble (waves crashing, ice breaking), and the loud, grating static of human activity (ships chugging, drilling machines). Scientists use special underwater microphones, called hydrophones, to record this concert. Their goal is to pick out specific "solos"—like a specific whale call—to understand if the ecosystem is healthy. But here's the problem: the human noise is so loud it drowns out the music, making it impossible to tell who is singing.

To fix this, researchers often try to use "noise reduction" tools, which act like a digital pair of noise-canceling headphones for the ocean. The idea is to scrub away the static so the animal sounds shine through. However, there's a catch. These tools change the sound in ways that might confuse the computer programs (AI) trained to recognize the animals. It's like if you cleaned a muddy painting so well that the colors changed, and the art expert no longer recognized the masterpiece. Before we trust these tools, we need a way to test them perfectly. We need a way to create fake ocean recordings where we know exactly how loud the music is and exactly how loud the noise is, so we can see if the cleaning tool helps or hurts.

This is exactly what Bram Cuyx and his team at the Flanders Marine Institute and KU Leuven set out to do. They realized that in the real world, it's nearly impossible to find a "clean" recording of an animal sound without any background noise to use as a perfect reference. So, instead of hunting for perfect recordings, they built a digital "sound lab." They created an open-source toolbox that lets anyone mix and match real animal sounds with real ocean noise to create thousands of brand-new, synthetic recordings. The magic of their system is that they know the "ground truth"—they know exactly which part of the fake recording is the animal and which part is the noise. They even generate a special "mask," like a stencil, that shows exactly where the animal's energy is hiding in the sound wave.

Using this toolbox, the team ran a series of experiments to see how well different AI models could find animal sounds when the ocean was noisy. They tested the models on recordings with Signal-to-Noise Ratios (SNR) ranging from -15 dB (extremely loud noise) to 20 dB (very clear sounds). They then applied a simple noise reduction filter, called a Wiener filter, to see if it made the AI smarter.

The results were a mix of surprises and warnings. First, they found that when the noise gets too loud (below -5 dB), the AI models essentially give up, performing no better than a random guess. This suggests that noise reduction is absolutely necessary for these tools to work in the real, messy ocean. However, the effect of the noise reduction wasn't the same for every model. For some AI models, like the one designed for whales (Google Whale), the noise reduction filter was a game-changer, significantly improving their ability to hear the signal. For others, like the model for reef species (SurfPerch), the filter didn't help much at all.

The study also showed that the simple Wiener filter they used had limits. At very low signal levels, the filter struggled to clean up the sound effectively. However, for the models that did improve, the filter helped boost their performance. The authors emphasize that this doesn't mean noise reduction is a magic bullet that solves everything. Instead, their findings suggest that we cannot just slap a noise filter onto any detection system and expect it to work. The relationship between the noise, the filter, and the specific AI model is complex.

In short, the paper introduces a vital new tool for scientists to build better, safer testing grounds for ocean listening technology. It proves that while cleaning up the noise can help, it must be done carefully and tested specifically for each AI model, because what helps one might not help another. The toolbox they created allows researchers to simulate these tricky scenarios without needing perfect real-world data, paving the way for more reliable ways to listen to the ocean's secrets.

Technical Summary: Acoustic Data Synthesis for Evaluating Noise Reduction in Bioacoustic Event Detection

Problem Statement
Passive acoustic monitoring (PAM) is increasingly used to assess underwater ecosystem health by analyzing soundscape events (biophony, geophony, and anthropophony). However, the utility of these recordings is often compromised by high levels of continuous noise from sources such as shipping, fishing, seismic surveys, and instrument artifacts. This noise results in low signal-to-noise ratios (SNR), hindering the ability of Sound Event Detection (SED) systems to disentangle and classify events.

While noise reduction (NR) front-ends are a common strategy in other fields like speech processing to enhance signal quality before analysis, their application in marine bioacoustics remains under-explored. A critical challenge is that NR algorithms alter the spectrotemporal characteristics of the input, thereby changing the input distribution for downstream embedding networks. Without controlled testing environments, it is difficult to determine if a pre-processing step genuinely improves detection or inadvertently degrades model performance. Furthermore, existing bioacoustic datasets typically lack "clean" ground truth labels because events are rarely recorded in isolation from their noisy environments, making it difficult to synthesize data with known SNR levels for rigorous evaluation.

Methodology
The authors propose an open-source data synthesis toolbox designed to generate marine acoustic recordings with fully controlled parameters (SNR, event count, duration, and noise origin). The methodology involves four key stages:

Event Presence Mask Generation:
To separate event energy from background noise in existing recordings, the authors utilize Non-Negative Matrix Factorization (NMF). They decompose the power spectrogram of a manually selected recording into basis spectra and temporal activations. By assuming the background noise is stationary, they estimate the noise power spectrum from the first second of the recording (where no event energy is present by design). A binary time-frequency mask is then generated by thresholding the residual power (reconstructed spectrogram minus estimated noise). This mask identifies the time-frequency bins containing event energy, allowing for the extraction of a "clean" event estimate.
Recording Synthesis:
The toolbox synthesizes new recordings by mixing the extracted "clean" event components with selected background noise fragments. The synthesis occurs in the time-frequency domain using the Short-Time Fourier Transform (STFT). The event component is scaled by a factor $\sigma$ to achieve a specific target SNR relative to the noise component. The process generates a synthetic signal $Y_{syn} = E_{syn} + N_{rec}$ , accompanied by a metadata file that explicitly details the source files, event onset/offset, the binary event mask, and the scaling factor used.
Noise Reduction Evaluation:
To demonstrate the toolbox's utility, the authors apply a simple Wiener filter as a pre-processing step. This filter operates in the time-frequency domain, estimating the event Power Spectral Density (PSD) recursively and applying a gain function based on the ratio of estimated event PSD to total signal PSD. The filter parameters (forgetting factor $\beta$ and minimum gain $\xi$ ) were tuned to balance noise reduction with the minimization of musical noise artifacts.
Experimental Setup:
The authors generated a dataset of 1,600 recordings (10 seconds each, 48 kHz) covering SNR levels from -15 dB to 20 dB, with 0 or 1 event per file. They evaluated five pre-trained bioacoustic models (Perch V2, SurfPerch, BirdNET, BioLingual, and Google Whale) by training linear classifiers on their embeddings. Performance was measured using F1-scores and ROC-AUC, comparing results on raw noisy data versus Wiener-filtered data.

Key Contributions

Open-Source Toolbox: The introduction of a tool capable of synthesizing bioacoustic datasets with precise control over SNR and event presence, accompanied by ground-truth binary time-frequency masks.
Evaluation Framework: A methodology to isolate and evaluate the impact of noise reduction pre-processing on SED pipelines, addressing the lack of "clean" ground truth in real-world bioacoustic data.
Empirical Analysis: A systematic investigation into how different pre-trained models respond to noise reduction at varying SNR levels, highlighting that the efficacy of pre-processing is model-dependent.

Results
The evaluation yielded several key findings:

SNR Dependency: All models exhibited a decline in performance (F1 and ROC-AUC) as SNR decreased. Notably, at SNR levels below -5 dB, all models performed at the level of random chance (ROC-AUC $\approx$ 0.5), indicating a threshold where current models fail to distinguish events from noise.
Model-Dependent Improvement: The impact of the Wiener filter was not uniform across models. The "Google Whale" model, which had the poorest baseline performance, showed the most significant improvement after noise reduction. Conversely, the "SurfPerch" model, which performed best overall, showed minimal gains from the filter.
Filter Limitations: The simple Wiener filter provided limited noise reduction at very low SNRs but showed increasing performance gains as the input SNR increased.

Significance and Claims
The paper asserts that the controlled synthesis of data is essential for isolating the effects of pre-processing steps in bioacoustic detection pipelines. The authors claim that their toolbox enables researchers to:

Rigorously test whether noise reduction strategies actually improve detection performance or merely alter the input distribution in ways that confuse models.
Augment existing datasets with new, controlled noise scenarios to improve model robustness.
Understand that the "best" noise reduction strategy is not universal but depends heavily on the specific architecture and training of the downstream detection model.

The authors conclude that while their specific implementation (a simple Wiener filter) showed promise, the framework allows for the future exploration of more advanced techniques, such as deep learning-based denoising or array processing, within a controlled evaluation environment. They emphasize that careful assessment of processing steps is necessary before integrating them into operational detection pipelines.

Technical Summary: Acoustic Data Synthesis for Evaluating Noise Reduction in Bioacoustic Event Detection

More like this