Improve automatic detection of animal call sequences with temporal context.

Area of Science:

• Bioacoustics
• Computational Biology
• Animal Communication

Background:

• Animals use complex songs for critical functions like mating and territory defense.
• Accurate recognition of animal vocalizations is vital for biological studies and conservation.
• Current methods may not fully capture the sequential nature of animal songs.

Purpose of the Study:

• To improve automatic recognition of animal songs by incorporating temporal context.
• To evaluate a hybrid deep learning model combining CNNs and LSTMs for song analysis.
• To assess the effectiveness of this model using fin whale vocalizations.

Main Methods:

• Developed a hybrid deep learning model integrating a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network.
• The CNN was trained to detect individual song notes (calls) in short audio segments.
• The LSTM processed sequential representations from the CNN to capture temporal patterns over longer time scales.
• The combined CNN+LSTM model was evaluated using recordings of fin whale songs.

Main Results:

• The CNN+LSTM models demonstrated improved performance compared to a baseline CNN model.
• Performance variance was reduced in the hybrid models.
• Area under the precision-recall curve increased by 9-17%.
• Peak F1-scores increased by 9-18%.

Conclusions:

• Integrating temporal information through hybrid CNN+LSTM models significantly enhances automatic recognition of animal songs.
• This approach offers a valuable method for improving the accuracy of wildlife sound analysis and transcription.
• The findings have direct implications for bioacoustic monitoring and conservation strategies.