Fine-Grained Bioacoustics Species Recognition via Multi-Feature Synergy and Ensemble Decision Optimization

Bioacoustics monitoring is vital for wildlife conservation, as over 60% of species rely on acoustic signals, yet manual surveys capture less than 40% of vocal activity. With endangered populations declining by 1–2% annually, the limitations of manual analysis being time-consuming, subjective, and prone to error necessitate scalable, automated solutions. This work proposes an Automated Animal Species Identification (AASI) system architecture using machine learning to monitor six target classes: lions, bears, dolphins, monkeys, donkeys, and elephants. The system utilizes a multi-spectral feature extraction framework to capture complex temporal and frequency-domain characteristics. Extracted features include Mel-frequency Cepstral Coefficients (MFCCs), Mel-Spectrograms, Chroma features, Spectral Contrast, Tonnetz, and various statistical descriptors like Zero-Crossing Rate (ZCR) and Spectral Centroid. This diverse feature set ensures robustness against background noise and overlapping calls. Following Exploratory Data Analysis (EDA) to assess class balance and feature distributions, the study evaluates traditional models such as Decision Tree, Nearest Centroid, and Gradient Boosting as baselines. To enhance performance, a Voting Ensemble Classifier (VEC) is proposed, integrating Support Vector Classifier (SVC) and Light Gradient Boosting Machine (LGBM). By employing a strategic voting mechanism, the AASI system achieves superior accuracy, stability, and generalization, providing a reliable framework for large-scale, real-time wildlife surveillance and biodiversity assessment.