Source-linked AI summary
Processing and classifying bird songs using wavelet techniques and supervised learning
Laura Lucia Dominguez Barrios, Fidel Aniano Causil Barrios, Alex Rodrigo dos Santos Sousa, Mariana Rodrigues Motta
TL;DR
Bird vocalizations recorded in noisy natural soundscapes are difficult to classify reliably. The study combines Bayesian wavelet shrinkage with acoustic feature extraction and supervised learning, finding stable species separation and informative low-order cepstral features.
Problem
Bird sounds degrade during transmission and recordings from natural soundscapes contain substantial environmental noise, motivating classification of invasive birds from bioacoustic signals.
Method
The framework applies Bayesian wavelet-based denoising with an Epanechnikov prior, then classifies recordings using MFCCs and spectral descriptors.
Results
Species groups remain well separated across m=4 and m=10, while lower cepstral components and selected spectral descriptors provide the most informative predictors.
Takeaways & Limitations
Bayesian wavelet shrinkage is presented as a computationally efficient preprocessing approach for recovering bird songs under high noise levels.
Takeaways & Limitations
The framework’s scalability to hyper-diverse soundscapes and its ability to detect rare or mimetic vocalizations in real-time surveillance remain future research directions.
Abstract
from arXiv · showhide
This study proposes an integrated framework for the processing and classification of invasive bird species vocalizations within natural soundscapes, characterized by high levels of environmental noise. We address the challenge of signal degradation by employing a Bayesian wavelet shrinkage methodology based on the Epanechnikov kernel prior, which offers a closed form decision rule and high computational efficiency for processing large bioacoustic datasets. The methodology was applied to recordings of three species obtained from the iNaturalist platform: \textit{Euphonia violacea}, \textit{Leiothrix lutea}, and \textit{Passer domesticus}. After signal denoising, we extracted a comprehensive set of features, including Mel-Frequency Cepstral Coefficients (MFCCs) and spectral indices such as entropy and zero-crossing rate. Several supervised learning models: Random Forest, Multinomial Logistic Regression and Support Vector Machine (SVM) were evaluated across different feature dimensionalities. Our results demonstrate that the proposed wavelet based preprocessing significantly enhances classification performance, with the SVM model achieving the highest accuracy (up to 0.9398) under a 10-dimensional MFCC configuration. This research provides a robust statistical tool for automated ecological monitoring and the management of biological invasions.
1 Introduction
The study targets bird-song classification in noisy natural soundscapes, combining Bayesian wavelet denoising with acoustic features and supervised learning.
- Natural soundscapes complicate species recognition through propagation degradation, variable source distance, vocal variation, mimicry, and recording biases.
- Wavelet methods are attractive because basis expansions represent signals sparsely, while Bayesian priors incorporate coefficient sparsity and support information.
- The framework removes silence, standardizes sampling, applies wavelet shrinkage, extracts MFCCs and related covariates, and compares supervised classifiers.
- The study evaluates whether this processing and classification framework can identify invasive birds from noisy bioacoustic recordings.
2 Wavelet-based estimation methodology
The methodology models noisy observations in a wavelet basis, estimates coefficients with an Epanechnikov-prior Bayesian shrinkage rule, and reconstructs denoised signals by inverse transformation.
- 2.1 Statistical model and the discrete wavelet transform: The discrete wavelet transform moves observations into the wavelet domain, and the inverse transform reconstructs the estimated function.
- 2.1 Statistical model and the discrete wavelet transform: The wavelet-domain model treats empirical coefficients as noisy versions of sparse unknown coefficients and estimates them coefficient by coefficient.
- 2.2 Bayesian wavelet shrinkage rule: The Epanechnikov prior combines a point mass at zero with a symmetric unimodal density, producing a posterior-mean shrinkage rule under squared loss.
- 2.3 Hyperparameters elicitation: Increasing alpha strengthens coefficient shrinkage, while resolution-level settings apply stronger shrinkage at finer scales where noise is concentrated.
- 2.2 Bayesian wavelet shrinkage rule: The rule is motivated by competitive performance in low signal-to-noise settings and uses an explicit decision rule for computationally efficient processing.
- 2.3 Hyperparameters elicitation: The hyperparameters beta and lambda adapt to coefficient magnitudes and fine-scale variability, allowing smoothing to respond to recording noise.
3.1 Euphonia violacea
Euphonia violacea is described as a colorful, melodious resident bird distributed across northeastern and eastern South America.
- Euphonia violacea inhabits forests, forest edges, parks, gardens, cocoa plantations, and citrus orchards across parts of South America.
- Its diet is primarily frugivorous, supplemented by nectar and insects when seasonally available.
- The species exhibits colorful plumage and a melodious song, with males, females, and juveniles differing in coloration.
3.2 Leiothrix lutea
Leiothrix lutea is presented as an introduced European species with documented establishment and a high assessed invasion risk based on biological, ecological, and socioeconomic factors.
- Leiothrix lutea became established in Europe through escapes or deliberate introductions, with established populations reported in France, Italy, Spain, and Portugal.
- 3.2.1 Risk analysis: The geographic-distribution figure summarizes 5,000 iNaturalist records of Leiothrix lutea observations worldwide.
- 3.2.1 Risk analysis: 81.5 is the final risk score for Leiothrix lutea, classifying it as presenting a very high invasion risk and supporting a recommendation for rejection.
- 3.2.1 Risk analysis: The assessment identifies strong establishment, survival, and spread capacity alongside biogeographical, socioeconomic, and invasion-enhancing risk factors.
3.3 Passer domesticus
This section characterizes Passer domesticus as a globally distributed species with high invasive risk and describes the data and acoustic features used for analysis.
- Risk analysis: 91.5 was the final Passer domesticus risk-assessment score, placing the species in the very high invasion-risk category.The assessment met the minimum criteria for a valid risk analysis.
- Data and features: Records for Euphonia violacea, Leiothrix lutea, and Passer domesticus were retrieved from iNaturalist and grouped by iconic taxon name.Recordings were queried using “song,” restricted to 2025, and records without valid audio were excluded.
- Acoustic features: The analysis considered MFCCs alongside entropy, zero-crossing rate, spectral centroid, RMS, and quartile-based descriptors.MFCCs represent spectral characteristics relevant to automated bird-species recognition, while the other measures describe spectral distribution, temporal crossings, energy, or dispersion.
3.4 Statistical analysis
The statistical analysis denoised recordings with Bayesian wavelet methods, extracted multiscale features, and compared supervised classifiers under variable acoustic conditions.
- Preprocessing: Bayesian wavelet shrinkage adaptively smoothed signals while preserving important temporal features before feature extraction.The denoising procedure shrank wavelet coefficients under a prior structure controlled by hyperparameters.
- Feature extraction: Wavelet decomposition produced energy-based detail-coefficient descriptors and signal standard deviations, forming a multiscale feature dataset.A Daubechies filter was applied, with total and mean detail-coefficient energy computed at each decomposition level.
- Evaluation design: The data were split into stratified training and testing subsets using 80% for training and 20% for testing.Stratification preserved class proportions across the two subsets.
- Classification: Random Forest, multinomial logistic regression, and kernel-based classifiers were trained and compared using cross-validation and test-set confusion matrices.Random Forest used 10-fold cross-validation, while logistic regression used 5-fold cross-validation; performance included overall and class-specific measures.
- Signal quality: SNR values varied substantially, with intermittent peaks and low-intensity periods reflecting heterogeneous recording conditions.Data 1 included several peaks exceeding 10 dB, whereas Data 2 had a narrower range concentrated at lower levels.
4 Results and discussion
The acoustic feature space separates the species, with the clearest clustering patterns differing between the two datasets. Variable-importance analyses show that discrimination relies on different combinations of cepstral and broader acoustic descriptors.
- PCA projection: Clear separation between species is observed in both datasets, with Data 2 forming more differentiated clusters and Data 1 showing slight overlap.Discrimination occurs mainly along the first principal component, while Data 2 exhibits minimal overlap and Data 1 shows some central overlap.
- Data 1 variable importance: MFCC-derived descriptors dominate discrimination between Passer domesticus and Leiothrix lutea across m = 4 and m = 10.Lower-order MFCC summaries, including minimum, mean, and median statistics, consistently show the highest importance values.
- Data 1 variable importance: MFCC min2 is the strongest predictor in the m = 4 Data 1 model, with MDA 14.56 and MDG 9.14.MFCC mean1, MFCC mean2, MFCC med1, and MFCC med2 also show high importance, linking discrimination to lower-order spectral-envelope structure.
- Data 1 variable importance: At m = 10, several higher-order MFCC descriptors show low or negative MDA values, indicating limited contribution and possible redundancy.The supplied analysis notes that adding cepstral dimensions can distribute explanatory power across correlated predictors and reduce interpretability.
- Data 2 variable importance: Data 2 discrimination is distributed across classical acoustic indices and cepstral variables rather than concentrated exclusively in low-order MFCCs.Entropy and ZCR remain influential, while lower-order MFCC descriptors retain moderate relevance and higher-order moments contribute additional information.
5 Final considerations
The study finds that Bayesian wavelet denoising combined with supervised learning supports accurate invasive-bird vocalization classification in noisy recordings. SVMs and higher-dimensional feature spaces performed best, while low-order MFCCs provided the strongest discrimination and broader scalability remains for future work.
- Bayesian wavelet shrinkage with an Epanechnikov prior recovered underlying bird songs from recordings with low signal-to-noise ratios using an explicit, computationally efficient rule.
- SVM consistently outperformed the other models, and expanding the feature space from 4 to 10 dimensions generally improved predictive metrics.The authors attribute SVM’s advantage to handling nonlinear species-specific acoustic boundaries.
- Lower-order MFCCs were the primary source of discrimination, although spectral entropy and zero-crossing rate also contributed to variable importance.The result links species identification mainly to stable low-frequency cepstral structures.
- The integrated denoising and supervised-learning pipeline offers a high-performance approach for automated monitoring of invasive birds.
- Future research should test scalability in hyper-diverse soundscapes and integrate deep-learning architectures for rare or mimetic vocalizations in real-time surveillance.