Noise-Resilient Bioacoustics Feature Extraction Methods and Their Implications on Audio Classification Performance:...

Area of Science:

• Bioacoustics
• Machine Learning
• Signal Processing

Background:

• Bioacoustics classification is vital for ecological surveillance and neonatal health monitoring.
• Environmental noise and signal variability challenge model reliability.
• Robust feature extraction and denoising are critical for accurate acoustic event interpretation.

Purpose of the Study:

• To systematically review advancements in noise-resilient feature extraction and denoising for bioacoustics classification.
• To explore methodological trends, model types, and cross-domain transferability.
• To assess evidence for real-world deployment in clinical and ecological settings.

Main Methods:

• Systematic review of 8 electronic databases up to December 2024.
• Inclusion of studies on audio-based classification using machine/deep learning with noise consideration.
• Data extraction and quality assessment by two independent reviewers using TRIPOD checklist.

Main Results:

• 132 studies met eligibility criteria, with deep learning and hybrid models dominating.
• Feature extraction was present in 96.2% of studies, commonly using Mel frequency cepstral coefficients and spectrograms.
• 47% of studies incorporated noise-resilient methods, but only 14.4% demonstrated real-world deployment.

Conclusions:

• Noise-resilient techniques enhance classification performance but face limited real-world application.
• Challenges include dataset heterogeneity, inconsistent reporting, and reliance on synthetic noise.
• Future research should focus on harmonized benchmarks, cross-domain generalization, and deployment strategies.