The stakes behind this computational problem are enormous. The Hainan gibbon (Nomascus hainanus) is critically endangered and endemic to China, surviving today only in Bawangling within Hainan Tropical Rainforest National Park. With such a restricted habitat and tiny population, continuous monitoring of survival and population dynamics is considered essential for maintaining regional biodiversity. Traditional visual surveys, however, are constrained by complex terrain, dense vegetation, adverse weather, and human interference, particularly in highly structured tropical rainforest ecosystems. Passive acoustic monitoring offers a non-invasive alternative: autonomous recording units collect animal sounds continuously over large scales and long periods, providing data for tracking population dynamics, investigating behavioral patterns, optimizing management, and assessing individual welfare. For a vocally active species like the gibbon, whose calls carry acoustic energy predominantly below 2 kHz and feature long durations and prominent harmonic structures, sound is the most practical window into its hidden lives.

Yet passive acoustic monitoring generates a flood of data, and efficiently extracting, annotating, and analyzing target calls from massive datasets remains a central bottleneck in ecoacoustics. Long-term continuous monitoring produces enormous volumes of recordings unrelated to research objectives, and pulling target signals out of them demands substantial expert time. During peak calling periods such as the dawn chorus, multiple species often vocalize simultaneously, creating high spectro-temporal overlap among sound sources that complicates source separation and single-species identification. Field recordings are also vulnerable to persistent biotic noise, abiotic noise from wind and rain, and anthropogenic disturbances. Such background noise can be closely coupled with the target call, masking critical acoustic features and rendering traditional analysis unreliable under low signal-to-noise ratio conditions.

Deep learning has delivered powerful tools for automated ecoacoustic analysis, including convolutional neural network architectures for bioacoustic source separation that have handled macaque, bottlenose dolphin, and Egyptian fruit bat calls, end-to-end models for bird song separation, and spectrogram transformation methods validated on bird species. But most of these approaches rely on supervised learning and require clean target signals as ground-truth references, which are often impractical to obtain in real-world ecological monitoring. Some architectures assume a fixed number of sound sources, an assumption that sits awkwardly with the dynamic nature of field soundscapes. Synthetic data simulation is a common workaround, but synthetic samples struggle to replicate the complex, non-stationary noise structures of natural environments, producing a domain mismatch that causes separation performance to decline sharply during field application. Unsupervised attempts have their own limits: some over-rely on synthetic training data, some are constrained by a fixed number of output sources and risk over-separation, some depend heavily on intermediate dataset quality, and some, built on speech enhancement models, carry an architectural bias toward human speech that can attenuate the high-frequency features of animal vocalizations. Moreover, existing studies have focused primarily on birds and marine mammals, leaving primate source separation largely unexplored.

The new framework builds on RemixIT, a teacher-student self-supervised approach for source separation that iteratively optimizes a student model using pseudo-targets generated by a teacher model, allowing training directly on mixed recordings without clean reference signals. The mechanism is elegant in its bootstrapping. A teacher model, trained on out-of-domain data, separates an input mixture into an estimated target call and an estimated background noise. The estimated noise is randomly permuted within each batch, then remixed with the estimated call to form a new bootstrapped mixture. The student model takes this bootstrapped mixture as input and learns to recover the teacher’s pseudo-targets under novel interference conditions, effectively teaching itself to extract the target call under ever-changing noise. An exponential moving average teacher, with a momentum coefficient of 0.01 following the original framework’s default configuration, keeps the two models synchronized throughout all experiments without further tuning. The separation backbone adopts the SuDoRM-RF architecture, which processes raw time-domain waveforms directly.

MC-RemixIT adds three purpose-built optimizations to this foundation. First, a lightweight call enhancement module sits at the front end of both models, integrating an attention mechanism to adaptively enhance discriminative features of target calls without significantly increasing computational load. The module maps the input waveform into a higher-dimensional feature space via a large-kernel 1D convolution, refines features through depthwise separable convolution blocks paired with lightweight Efficient Channel Attention, and maps features back to a single-channel waveform combined with the original signal through a residual connection, enabling explicit noise suppression while retaining informative characteristics of the original waveform. Second, a signal cross-correlation constraint, computed in the frequency domain and penalizing the normalized peaks of the complete cross-correlation sequence, minimizes the statistical correlation between the estimated call and the estimated background noise, pushing the two sources toward greater acoustic independence. Third, a multi-constrained loss for student training combines the primary separation objective with spectral flatness and energy consistency terms, keeping the total energy of separated signals consistent with the input mixture and preserving meaningful spectral characteristics while curbing unnecessary spectral fluctuations.

The experiments drew on two previously published Hainan gibbon recording campaigns, both collected in Bawangling. One dataset, gathered mainly from September 2021 to December 2022 with daily continuous recordings from 06:00 to 18:00 at a 48 kHz sampling rate, supplied 500 call segments with relatively low background noise for teacher training; these were mixed with randomly selected human activity noise from the WHAM! Noise dataset at varying signal-to-noise ratios to generate 4,000 mixed samples. The second campaign, collected between March and August 2016 with devices recording eight hours from sunrise each day at mostly 9,600 Hz, supplied the student training data, which contains substantially more field interference; some samples were randomly paired and mixed to simulate overlapping calls from multiple individuals. All recordings were converted to mono, uniformly resampled to 16 kHz, a rate deemed sufficient to preserve the calls’ essential low-frequency characteristics, and segmented with a 10-second sliding window and 1-second hop size. Critically, dataset partitioning was performed at the recording level rather than the clip level, keeping all segments from the same original recording within the same subset to prevent information leakage. The final teacher dataset contained 4,000 training, 500 validation, and 500 test samples; the student dataset contained 10,000 training, 1,000 validation, and 1,000 test samples.

Because field recordings lack clean references, traditional reference-based metrics such as SI-SDR and PESQ could not be applied, so the team evaluated separation using DNSMOS, a non-intrusive perceptual quality assessment that scores signals from 1 to 5 along three dimensions: Speech Quality, Background Noise Quality, and Overall Quality. On representative low-SNR field recordings, MC-RemixIT’s call estimates showed clearer harmonic structures, more continuous syllable contours, and less residual background interference than the baseline RemixIT in spectrogram comparisons, with the baseline’s noise estimates containing noticeable residual call components that were substantially reduced by the new method. Quantitatively, the Speech Quality score rose from 2.82 to 2.93, Background Noise Quality from 2.94 to 3.06, and Overall Quality from 2.57 to 2.63 on those samples. Across the broader dataset, the mean Background Noise Quality score increased from 2.03 to 2.32 with reduced standard deviation, indicating more stable noise suppression. Comparison with Biodenoising, a state-of-the-art bioacoustic denoising approach, proved revealing: Biodenoising achieved slightly better noise suppression with a mean Background Noise Quality of 2.35, but suffered substantial signal quality degradation, with a mean Speech Quality of 2.51 and Overall Quality of 1.94, both significantly lower than MC-RemixIT. An ablation study confirmed each module’s contribution, with the full framework lifting Overall Quality from 2.44 to 2.54 and Background Noise Quality from 2.03 to 2.32 while holding Speech Quality near 3.0.

To test whether the separated signals retain task-relevant information, the researchers constructed a downstream gibbon call presence detection task, applying source separation exclusively to the testing subset while training detection models only on preprocessed audio without separation. Four convolutional architectures were evaluated under three input conditions: raw preprocessed segments, RemixIT estimates, and MC-RemixIT estimates. On the overall test set, MC-RemixIT improved detection across all models: ResNet50 accuracy rose from 80.60% to 90.80% with an F1-score of 91.31%, ResNet34 from 81.84% to 88.31%, and MobileNetV3s from 76.62% to 88.56%, a larger relative gain for the lightweight network. On the Dufourq test subset, MC-RemixIT achieved the highest precision, recall, and F1-score across all four models, with ResNet50 reaching 91.93%, 91.63%, and 91.60%, gains of 8.24, 7.66, and 6.40 percentage points in F1, recall, and precision over the unseparated baseline. On the much smaller Wang test subset, which contained only 47 samples and no absence class, the relative performance of the three conditions varied across models, a variation the authors attribute to the subset’s size, class composition, and recording-level partitioning, and they note that further evaluation with larger, more class-balanced datasets would help fully assess applicability across recording sources.

The study is a computational evaluation on field monitoring recordings, not a demonstration of conservation outcomes, and the authors are candid about its limits. DNSMOS was originally developed for human speech enhancement and may not fully capture species-specific acoustic characteristics, which is why the downstream detection task serves as a complementary, task-oriented check. Separation performance may decline under extremely low-SNR conditions where calls are heavily masked, and the framework focuses on separating target calls from background noise rather than disentangling overlapping calls from multiple individuals, an individual-level challenge that remains open. Even so, the findings indicate that self-supervised source separation can serve as an effective front-end preprocessing step for automated ecoacoustic analysis, reducing reliance on clean reference recordings that are nearly impossible to obtain in the field. Because the approach was designed specifically for a species facing severe data constraints, the authors suggest the optimization strategy may offer a useful reference for other vocally active species monitored under similar conditions, potentially improving the quality of acoustic data available for ecological analysis and conservation research far beyond one rainforest in Hainan.

Subject of Research: Self-supervised AI source separation of Hainan gibbon calls in passive acoustic monitoring

Article Title: A Self‐Supervised Approach for Hainan Gibbon Call Separation in Passive Acoustic Monitoring

Article References: Zi, J., Xu, N., Dong, Y., Xia, Y., Wang, W., & Chen, F. (2026). A Self‐Supervised Approach for Hainan Gibbon Call Separation in Passive Acoustic Monitoring. Ecology and Evolution, 16(10), Article e74542. https://doi.org/10.1002/ece3.74542

Image Credits: AI Generated

DOI: 10.1002/ece3.74542

Keywords: Hainan gibbon, passive acoustic monitoring, self-supervised learning, source separation, bioacoustics, deep learning, MC-RemixIT, DNSMOS, conservation technology, tropical rainforest soundscapes, teacher-student framework, call detection