Dual Task Deep Learning Framework for Automatic Recognition of Avian Species and Song Identification

• Anju L,
• Ganesh Vaidyanathan S,

Vol. 20, No. 1, January 31, 2026
10.3837/tiis.2026.01.026, Download Paper (Free):

• adaptive quantization
• Attention-based Deep Belief Network
• Avian Vocalization Analysis
• Bioacoustics
• Bird Song Detection
• Deep Feature Extraction
• Stacked Gated Recurrent Units.

Abstract

Accurate identification of bird species and their songs is essential for avian biodiversity research, ecological monitoring, and conservation. However, traditional methods often struggle with limited feature representations and fail to capture complex temporal and spectral patterns in bird vocalizations. This paper proposes a Dual-Task Deep Learning (DTDL) framework for simultaneous bird species and song identification. The system utilizes a curated dataset comprising recordings from seven bird species, each with three distinct vocalizations. Preprocessing involves noise filtering, silence removal, and signal normalization. A novel Adaptive Two-Region Piecewise Uniform Quantizer (ATR-PWUQ), driven by a Deep Neural Network (DNN), adaptively adjusts decision thresholds to enhance signal representation. Feature extraction combines handcrafted audio features (e.g., MFCCs, time-frequency, harmonic) and deep features derived from Mel-spectrograms using the SqueezeNet model. These features are then classified using a hybrid Bidirectional Long Short-Term Memory with Stacked Gated Recurrent Units (Bi-LSTM-SGRU) for species identification and an Attention-based Deep Belief Network (A-DBN) for song classification. The Bi-LSTM-SGRU model achieves a bird species identification accuracy of 98.11%, while the A-DBN model achieves a song recognition accuracy of 98.98%. The system significantly outperforms existing models across all evaluation metrics. The proposed DTDL framework demonstrates high robustness and precision in avian acoustic analysis, making it a promising tool for real-world ecological applications and automated biodiversity monitoring.

Statistics

Cite this article