Separation of overlapping sources in bioacoustic mixtures

Semiautomated generation of species-specific training data from large, unlabeled acoustic datasets for deep supervised birdsong isolation

Background

Bioacoustic monitoring is an effective and minimally invasive method to study wildlife ecology. However, even the state-of-the-art techniques for analyzing birdsongs decrease in accuracy in the presence of extraneous signals such as anthropogenic noise and vocalizations of non-target species. Deep supervised source separation (DSSS) algorithms have been shown to effectively separate mixtures of animal vocalizations. However, in practice, recording sites also have site-specific variations and unique background audio that need to be removed, warranting the need for site-specific data.

Methods

Here, we test the potential of training DSSS models on site-specific bird vocalizations and background audio. We used a semiautomated workflow using deep supervised classification and statistical cleaning to label and generate a site-specific source separation dataset by mixing birdsongs and background audio segments. Then, we trained a deep supervised source separation (DSSS) model with this generated dataset. Because most data is passively-recorded and consequently noisy, the true isolated birdsongs are unavailable which makes evaluation challenging. Therefore, in addition to using traditional source separation (SS) metrics, we also show the effectiveness of our site-specific approach using metrics commonly used in ornithological analyses such as automated feature labeling and species-specific trilateration accuracy.

Results

Our approach of training on site-specific data boosts the source-to-distortion, source-to-interference, and source-to-artifact ratios (SDR, SIR, and SAR) by 9.33 dB, 24.07 dB, and 3.60 dB respectively. We also find our approach allows for automated feature labeling with single-digit mean absolute percent error and birdsong trilateration accuracy with a mean simulated trilateration error of 2.58 m.

Conclusion

Overall, we show that site-specific DSSS is a promising upstream solution for wildlife audio analysis tools that break down in the presence of background noise. By training on site-specific data, our method is robust to unique, site-specific interference that caused previous methods to fail.

Source separation with an acoustic vector sensor for terrestrial bioacoustics

Passive acoustic monitoring is emerging as a low-cost, non-invasive methodology for automated species-level population surveys. However, systems for automating the detection and classification of vocalizations in complex soundscapes are significantly hindered by the overlap of calls and environmental noise. We propose addressing this challenge by utilizing an acoustic vector sensor to separate contributions from different sound sources. More specifically, we describe and implement an analytical pipeline consisting of (1) calculating direction-of-arrival, (2) decomposing the azimuth estimates into angular distributions for individual sources, and (3) numerically reconstructing source signals. Using both simulation and experimental recordings, we evaluate the accuracy of direction-of-arrival estimation through the active intensity method (AIM) against the baselines of white noise gain constraint beamforming (WNC) and multiple signal classification (MUSIC). Additionally, we demonstrate and compare source signal reconstruction with simple angular thresholding and a wrapped Gaussian mixture model. Overall, we show that AIM achieves higher performance than WNC and MUSIC, with a mean angular error of about 5°, robustness to environmental noise, flexible representation of multiple sources, and high fidelity in source signal reconstructions.

BioCPPNet: automatic bioacoustic source separation with deep neural networks

We introduce the Bioacoustic Cocktail Party Problem Network (BioCPPNet), a lightweight, modular, and robust U-Net-based machine learning architecture optimized for bioacoustic source separation across diverse biological taxa. Employing learnable or handcrafted encoders, BioCPPNet operates directly on the raw acoustic mixture waveform containing overlapping vocalizations and separates the input waveform into estimates corresponding to the sources in the mixture. Predictions are compared to the reference ground truth waveforms by searching over the space of (output, target) source order permutations, and we train using an objective function motivated by perceptual audio quality. We apply BioCPPNet to several species with unique vocal behavior, including macaques, bottlenose dolphins, and Egyptian fruit bats, and we evaluate reconstruction quality of separated waveforms using the scale-invariant signal-to-distortion ratio (SI-SDR) and downstream identity classification accuracy. We consider mixtures with two or three concurrent conspecific vocalizers, and we examine separation performance in open and closed speaker scenarios. To our knowledge, this paper redefines the state-of-the-art in end-to-end single-channel bioacoustic source separation in a permutation-invariant regime across a heterogeneous set of non-human species. This study serves as a major step toward the deployment of bioacoustic source separation systems for processing substantial volumes of previously unusable data containing overlapping bioacoustic signals.