Source-linked AI summary
2-Dimensional spectral gating for denoising bioacoustics recordings
Julien Boussard, Mélisande Teng, Sulagna Saha, Mario Gallego-Abenza
TL;DR
Open-environment variation in noise and vocalizations complicates robust bioacoustic analysis, while frequency-independent spectral gating does not use cross-frequency signal structure. The paper extends Noisereduce with a two-dimensional, Gaussian-smoothed spectrogram mask. It reports improved denoising on the birdsong benchmark and qualitatively on sperm-whale and flycatcher recordings, while noting that downstream-task improvements remain unevaluated.
Problem
Open-environment recordings vary in noise, signal-to-noise ratio, and vocalizations, challenging robust ecological analyses; frequency-independent gating does not use signal structure across frequencies.
Source: §1 · p. 1Method
Results
2D-SG outperforms evaluated methods across the birdsong benchmark metrics except for a nonsignificant SI-SDR difference at 0 dB, and qualitatively improves denoising on whale and flycatcher examples.
Source: §4.1 · pp. 3–4, §4.2 · p. 4Takeaways & Limitations
The method offers a simple denoising option for bird and marine-mammal recordings without machine-learning expertise or additional data collection.
Source: §1 · p. 1, §4.2 · p. 4The method adds a Gaussian-window hyperparameter that may need manual tuning, and its effects on downstream detection or sound-event classification have not been evaluated.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Isolating vocalizations from noise in bioacoustics recordings is a prerequisite to many ecological analyses, including species identification, animal communication understanding, and individual or population-level variability studies. However, when recordings are acquired in open environments, vocalizations, noise, or signal-to-noise ratio can vary widely across individuals, species, environment, and recording conditions, making it hard to develop robust and generalizable methods for ecological analyses. To account for these challenges, noise reduction techniques are used to remove noise before downstream analyses. Popular methods such as Noisereduce rely on spectral gating, which estimates a noise threshold for each frequency channel. We propose a further improvement to Noisereduce, leveraging the fact that most animal sounds have structure across multiple frequencies. We apply our method on bird and marine mammals recordings and show that our extension of Noisereduce leads to improved denoising in both above and under water acoustic recordings, without impacting speed of preprocessing.
1. INTRODUCTION
Open-environment recordings vary in noise, signal-to-noise ratio, and vocalizations, challenging ecological analyses and robust generalization. The paper extends Noisereduce to exploit vocalization structure across frequencies with a simple method that requires no training data or machine-learning expertise.
1. INTRODUCTION
- Open-environment variation in noise, signal-to-noise ratio, and vocalizations complicates accurate extraction for ecological analyses and robust generalization across recording conditions.The paper also notes that machine-learning approaches can overfit to background noise rather than extract useful information.Source: §1 · p. 1
- Conventional filters are fast and efficient but generally assume additive Gaussian noise and generalize poorly to natural noise types.The Wiener filter assumes stationary signal and additive noise; the Savitzky-Golay filter fits a low-degree polynomial to a sliding window.Source: §1 · p. 1
- Noisereduce estimates a threshold independently for each frequency band, so it does not leverage signal structure spanning multiple frequencies.It masks values below each band’s mean plus a z-score-scaled standard deviation.Source: §1 · p. 1
- Deep-learning denoisers may require large datasets, expertise, labels, or a good noise model, which can hinder adoption and generalization.The paper does not compare 2D-SG with deep-learning methods because it targets a simple, efficient approach applicable without additional labeling or data collection.Source: §1 · p. 1
- 2D-SG extends Noisereduce by leveraging animal vocalizations’ structure across multiple frequencies, without requiring machine-learning expertise or additional data labeling or collection.The authors do not compare it with deep-learning methods.Source: §1 · p. 1
2. METHOD
2D-SG extends spectral gating by basing its mask on a Gaussian-smoothed, two-dimensional spectrogram rather than the raw spectrogram. It estimates noise thresholds by frequency channel, applies and optionally smooths the mask, then reconstructs the denoised recording.
2. METHOD
- 2D-SG computes a Gaussian-smoothed local average of the spectrogram, estimates per-channel noise thresholds, masks values below threshold, and inverts the masked transform to reconstruct audio.The illustrated pipeline converts raw audio to a spectrogram, smooths it, estimates a threshold, applies a mask, and reconstructs the signal.Source: §2 · p. 2, Figure 1 · p. 2
- The Gaussian window can be wider along frequency for short vocalizations, while optional mask smoothing helps avoid spectral leakage, discontinuity, and sampling artifacts.The paper names the Gibbs phenomenon as one such artifact.Source: §2 · p. 2
- Unlike Noisereduce, 2D-SG computes its mask from the spectrogram’s 2D local average, retaining signal where it is present across multiple frequencies.Its noise threshold can still vary over time to adapt to changing noise levels.Source: §2 · p. 2
3. EXPERIMENTS
The evaluation combines a quantitative birdsong benchmark with qualitative recordings of sperm whales and pied flycatchers. Methods are compared using four denoising metrics, with key parameters tuned on a selected recording for each SNR level and noise type.
3.1. Datasets
- The birdsong NOIZEUS benchmark contains 70 forty-second recordings from 14 European starlings, mixed with eight soundscape noise categories at four SNR levels.The source recordings were made in an acoustically isolated chamber, and the added noise enables comparison with clean recordings.Source: §3.1 · p. 3
- Real-world qualitative evaluation uses Mediterranean sperm-whale clicks and European pied-flycatcher songs, representing vocalizations relevant to echolocation, communication, and repertoire studies.The dataset descriptions emphasize click detection and comparisons of vocalizations across individuals and populations.Source: §3.1 · p. 3
3.2. Evaluation
- The benchmark compares 2D-SG with Noisereduce, Wiener, and Savitzky-Golay filters using SegSNR, SI-SDR, LSD, and MCD.The metrics assess noise removal, reconstruction quality, full-spectrum reconstruction, and spectral-envelope reconstruction; significance versus Noisereduce is tested with an independent two-sample t-test at 95% confidence.Source: §3.2 · p. 3
- For each SNR level and noise type, the main method parameter is selected on one B335 recording to maximize SI-SDR, then applied to the other recordings.For real-world recordings, the noise threshold is manually tuned for visually best spectrograms; the paper reports that the selected parameter can then be used across the dataset without adding preprocessing time.Source: §3.2 · p. 3
4. RESULTS
On the birdsong NOIZEUS benchmark, 2D-SG outperforms the evaluated alternatives across the reported metrics, with a stated nonsignificant exception for SI-SDR at 0 dB. Qualitative examples also show improved denoising on birdsong, sperm-whale, and flycatcher recordings.
4.1. Birdsong NOIZEUS dataset
- 2D-SG outperforms the other evaluated methods on all four birdsong NOIZEUS metrics, except SI-SDR at 0 dB, where the difference is not statistically significant.The comparison reports Seg-SNR, SI-SDR, MCD, and LSD across SNR levels; the table caption identifies the methods and metric layout.Source: §4.1 · p. 3, Table 1 · p. 3
- Across all SNR levels, 2D-SG yields lower LSD and MCD; differences are statistically significant for MCD at SNR 5 and 15 recordings and LSD at SNR 5, 10, and 15 recordings.Source: §4.1 · p. 4
- For the B335 town-noise example, 2D-SG removes more noise around 15–20 kHz and preserves more signal above 20 kHz and around 3.5 seconds.Figure 2 presents raw, Noisereduce, and 2D-SG spectrograms with shared colorbars within each column.Source: §4.1 · p. 4, Figure 2 · p. 4
4.2. Qualitative analysis on real-world recordings
- On the sperm-whale and pied-flycatcher examples, 2D-SG qualitatively removes more noise while better preserving vocalizations than Noisereduce.The whale example makes a fourth click near 16 seconds more apparent; the flycatcher example is one song.Source: §4.2 · p. 4
5. DISCUSSION
2D-SG needs tuning of its Gaussian-window size and z-score threshold, and its performance in downstream tasks remains unevaluated. The authors note it may be less effective than supervised denoising when labeled data are available, especially at low SNR.
5. DISCUSSION
- 2D-SG adds a Gaussian-window hyperparameter, and its default settings may require manual tuning alongside the z-score threshold.The authors say defaults may not be optimal across vocalization types; an adaptive noise threshold is a possible extension.Source: §5 · p. 4
- With labeled data, 2D-SG may not outperform supervised denoising, particularly for low-SNR recordings.The authors also note that 2D-SG requires no machine-learning expertise or training.Source: §5 · p. 4
- Downstream detection and sound-event classification improvements remain unevaluated.The authors describe denoising results as promising but have not tested these downstream tasks.Source: §5 · p. 4