A Novel Bird Sound Recognition Method Based on Multifeature Fusion and a Transformer Encoder.

Area of Science:

• Ornithology
• Bioacoustics
• Artificial Intelligence

Background:

• Accurate bird identification is essential for ecosystem and biodiversity studies.
• Previous bird sound recognition methods lacked accuracy due to reliance on single features and ignoring spatial information.

Purpose of the Study:

• To develop a robust bird sound recognition method using advanced deep learning techniques.
• To enhance the accuracy of bird identification for improved biodiversity monitoring and conservation.

Main Methods:

• Employed multiple convolutional neural networks and a transformer encoder for bird sound analysis.
• Utilized feature fusion to combine multiple acoustic features (MFCC, Chroma, Tonnetz) for comprehensive representation.
• Leveraged a transformer encoder to capture positional relationships within bird sound features.

Main Results:

• Achieved 97.99% accuracy, 96.14% recall, 96.88% F1 score, and 97.97% precision on the Birdsdata dataset.
• Attained 93.18% accuracy, 92.43% recall, 93.14% F1 score, and 93.25% precision on the Cornell Bird Challenge 2020 dataset.
• Demonstrated superior performance compared to previous methods by integrating multiple features and spatial information.

Conclusions:

• The proposed method offers a reliable and accurate solution for bird sound identification.
• This advancement supports more effective biodiversity monitoring and conservation strategies through precise bird classification.
• The integration of multiple neural networks and transformer encoders shows significant potential for bioacoustic analysis.