A REVIEW OF

AUTOMATED BIRD SOUND

RECOGNITION & ANALYSIS

IN THE

NEW AI ERA

(8th SLAAI-International Conference on Artificial Intelligence – SLAAI-ICAI-2024)

Thompson, M. (2023) Bird song visualized slides. https://academy.allaboutbirds.org/bird-song-visualized-slides/.

“Bird sounds provide key ecological insights”

Introduction

• Prerequisite to a research on audio augmentation strategies for improving learning-based birdsong classification.

​

• Automated recognition of bird sounds plays a crucial role in monitoring and conserving biodiversity

Objectives of the Study

Introduction

• Current State Analysis
• Review recent advancements in bird sound recognition and classification (2021–2023).
• Identify strengths, limitations, and challenges in existing approaches.

​

• Guidance for Future Research
• Highlight key trends and emerging techniques in machine learning for bird sounds.

​

• Guidance for Resource-Limited Regions
• Explore cost-effective and accessible solutions tailored to underdeveloped regions.

1. Challenges in Bird Sound Recognition

• Data quality and noise
• Bird sounds are often recorded in noisy environments.
• Class imbalance
• A common problem in biodiversity datasets
• Domain shifts in recordings
• Sounds from one region might not match those from another
• Computational constraints
• A barrier in resource-limited settings

Data Quality and Noise

• Sources of noise: environmental, overlapping calls

​

• Noise reduction techniques:
• Spectral subtraction
• Wiener filtering
• Denoising autoencoders
• GANs

​

1. Challenges in Bird Sound Recognition

Priyadarshani, N. et al. (2016) 'Birdsong Denoising Using Wavelets,' PLoS ONE, 11(1), p. e0146790. https://doi.org/10.1371/journal.pone.0146790.

Class Imbalance and Domain Shifts

• Class imbalance: Overrepresentation of common species
• Solutions:
• Oversampling
• Data augmentation
• Class weighting

​

• Domain shift: Variability across regions/habitats
• Solutions:
• Domain adaptation
• Fine-tuning for specific environments

​

1. Challenges in Bird Sound Recognition

2. Efficient Data Collection and Preprocessing

• Passive Acoustic Monitoring for large-scale data

​

• Noise reduction
• Spectral subtraction, Wiener filtering

​

• Spectrograms
• STFT, MFCCs

2. Efficient Data Collection and Preprocessing

Passive Acoustic Monitoring (PAM)

​

• Advantages:
• Continuous monitoring
• Large-scale data collection
• Cost-effective

​

• Challenges:
• Data storage
• Noise interference

​

AudioMoth low-cost, open-source acoustic sensor ©openacousticdevices.info

​

• Traditional:
• Spectral subtraction
• Wiener filtering

​

​

​

• Advanced:
• Denoising autoencoders
• GANs (e.g., DVAD model)

​

Noise Reduction Techniques

2. Efficient Data Collection and Preprocessing

Zhang, C. et al. (2024) 'Automatic bioacoustics noise reduction method based on a deep feature loss network,' Ecological Informatics, 80, p. 102517. https://doi.org/10.1016/j.ecoinf.2024.102517.

Noise Reduction Techniques

2. Efficient Data Collection and Preprocessing

Conditional Latent Diffusion architecture

Gibbons, A. et al. (2024) 'Generative AI-based data augmentation for improved bioacoustic classification in noisy environments,' arXiv (Cornell University) [Preprint]. https://doi.org/10.48550/arxiv.2412.01530.

​

2. Efficient Data Collection and Preprocessing

Spectrograms and Feature Extraction

• Spectrograms
• STFT
• MFCCs

​

• Feature extraction
• Handcrafted vs. Learned Features

Unraveling Bird Sounds with Spectrograms (no date). https://c2s2.engineering.cornell.edu/blogposts/FA23/UnravelingBirdSoundsWithSpectrograms.

​

3. Deep Learning Models for Bird Sound Recognition

• CNN for spatial patterns
• RNN for temporal patterns
• Hybrid models for enhanced accuracy

3. Deep Learning Models for Bird Sound Recognition

A schematic of hybrid models for classification.

Gupta, G. et al. (2021) 'Comparing recurrent convolutional neural networks for large scale bird species classification,' Scientific Reports, 11(1). https://doi.org/10.1038/s41598-021-96446-w.

3. Deep Learning Models for Bird Sound Recognition

​

• Strengths:
• Spatial feature extraction
• High accuracy

​

​

​

• Weaknesses:
• Requires large datasets
• Computationally expensive.

CNNs: Capturing Spatial Patterns

Xie, J. et al. (2019) 'Investigation of Different CNN-Based Models for Improved Bird Sound Classification,' IEEE Access, 7, pp. 175353–175361. https://doi.org/10.1109/access.2019.2957572.

​

RNNs: Capturing Temporal Dependencies

• Strengths:
• Models sequence and timing of bird sounds

​

• Weaknesses
• Struggles with long sequences
• Computational cost

​

3. Deep Learning Models for Bird Sound Recognition

​

Hybrid Models and Innovations

• CNN-RNN hybrids for spatial and temporal analysis

​

• Innovations
• Attention mechanisms
• Transformers

​

• Real-world application examples

3. Deep Learning Models for Bird Sound Recognition

E, D. et al. (2022) 'Artificial Humming Bird Optimization–Based Hybrid CNN-RNN for Accurate Exudate Classification from Fundus Images,' Journal of Digital Imaging, 36(1), pp. 59–72. https://doi.org/10.1007/s10278-022-00707-7.

4. Efficient Techniques for Resource-Limited Settings

• Data augmentation to balance datasets
• Transfer learning for small datasets
• Lightweight models like MobileNet

Data Augmentation Techniques

• Examples:
• Pitch shifting
• Time-stretching
• Noise addition

​

• Pros: Increases dataset diversity
• Cons: Risk of introducing artifacts

​

4. Efficient Techniques for Resource-Limited Settings

Transfer Learning

• Pre-trained models:
• Leveraging large datasets

​

• Fine-tuning for specific bird sound tasks

​

4. Efficient Techniques for Resource-Limited Settings

Transfer Learning

4. Efficient Techniques for Resource-Limited Settings

Kumar, S.V.S. and Kondaveeti, H.K. (2024) 'Bird species recognition using transfer learning with a hybrid hyperparameter optimization scheme (HHOS),' Ecological Informatics, 80, p. 102510. https://doi.org/10.1016/j.ecoinf.2024.102510.

​

Transfer Learning

4. Efficient Techniques for Resource-Limited Settings

Miyaguchi, A. et al. (2024) 'Transfer Learning with Pseudo Multi-Label Birdcall Classification for DS@GT BirdCLEF 2024,' arXiv (Cornell University) [Preprint]. https://doi.org/10.48550/arxiv.2407.06291.

​

• Examples:
• MobileNet
• TinyML

​

• Strengths:
• Reduced computation
• Real-time capability

​

• Use cases:
• Field sensors
• Mobile apps

Lightweight Models for Efficiency

4. Efficient Techniques for Resource-Limited Settings

Huang, Z. et al. (2024) 'TinyChirp: Bird Song Recognition Using TinyML Models on Low-power Wireless Acoustic Sensors,' arXiv (Cornell University) [Preprint]. https://doi.org/10.48550/arxiv.2407.21453.

​

5. Real-World Applications and Success Stories

BirdCLEF 2024

5. Real-World Applications and Success Stories

Permana, S.D.H. et al. (2021) 'Classification of bird sounds as an early warning method of forest fires using Convolutional Neural Network (CNN) algorithm,' Journal of King Saud University - Computer and Information Sciences, 34(7), pp. 4345–4357. https://doi.org/10.1016/j.jksuci.2021.04.013.

​

xeno-canto :: Sharing wildlife sounds from around the world (no date). https://xeno-canto.org/.

​

AudioMoth | Open Acoustic Devices (no date). https://www.openacousticdevices.info/audiomoth.

​

6. Future Directions

• Advanced Noise Reduction
• Incorporating adaptive filters and self-supervised learning

​

• Real-Time and Scalable Solutions
• Developing ultra-efficient edge computing models for on-site analysis
• Integrating IoT-enabled sensors for large-scale deployment
• Enabling automated species tracking and anomaly detection

​

• Ethical and Responsible AI in Conservation

Thank you!

Corresponding Author:

​

Akila Maithripala

kil@ucsc.cmb.ac.lk

Automatic vocalisation detection delivers reliable, multi-faceted, and global avian biodiversity monitoring

DOI: 10.1101/2023.09.14.557670

License: CC BY-NC-ND 4.0

​

Supplementary