Automatic recognition of giant panda vocalizations using wide spectrum features and deep neural network

Area of Science:

• Bioacoustics
• Artificial Intelligence
• Animal Behavior

Background:

• Giant pandas (GPs) possess a complex vocal repertoire crucial for their social interactions and behavior.
• Accurate classification of GP vocalizations is essential for ecological monitoring and conservation efforts.
• Existing automated systems often lack the necessary feature extraction capabilities for nuanced bioacoustic analysis.

Purpose of the Study:

• To develop and validate an automated system for recognizing giant panda vocalizations.
• To introduce a novel deep neural network (DNN) architecture for enhanced bioacoustic feature extraction.
• To achieve high accuracy in classifying diverse giant panda vocalization categories.

Main Methods:

• Recorded and labeled over 12,800 giant panda vocal samples from the Chengdu Research Base of Giant Panda Breeding.
• Developed a novel deep neural network (DNN) named 3Fbank-GRU.
• Extracted acoustic features using Mel filter bank (Fbank), Medium Mel Filter bank (MFbank), and Reversed Mel Filter bank (RFbank) to capture low, medium, and high frequencies.

Main Results:

• The 3Fbank-GRU model achieved over 95% accuracy in recognizing 16 distinct giant panda vocalization categories.
• The system demonstrated superior performance compared to traditional methods by incorporating multi-frequency band features.
• The proposed method enables accurate labeling of large-scale datasets of giant panda vocalizations.

Conclusions:

• The developed automatic vocalization recognition system is highly accurate and effective for giant panda research.
• The 3Fbank-GRU model offers a significant advancement in bioacoustic analysis for endangered species.
• This technology can be widely applied for monitoring giant panda populations and understanding their behavior through passive acoustic recording.