An Efficient Time-Domain End-to-End Single-Channel Bird Sound Separation Network

Area of Science:

• Bioacoustics
• Machine Learning
• Computational Biology

Background:

• Bird vocalizations are crucial for species identification and communication.
• Field recordings often contain mixed bird sounds, complicating analysis.
• Accurate bird sound separation is essential for ecological research and species monitoring.

Purpose of the Study:

• To develop an efficient end-to-end network for separating mixed bird sounds.
• To reduce computational resources and improve separation speed.
• To facilitate automatic bird species identification on mobile and edge devices.

Main Methods:

• Proposed a novel dual-path tiny transformer network for time-domain bird sound separation.
• Employed a supervised learning framework for training the separation network.
• Focused on reducing network parameters and floating-point operations for enhanced efficiency.

Main Results:

• Achieved significant separation performance with SI-SNRi of 19.3 dB and SDRi of 20.1 dB.
• Demonstrated substantially reduced parameters and computations compared to DPRNN and DPTNet.
• Exhibited higher separation efficiency and faster processing speeds.

Conclusions:

• The proposed network offers a computationally efficient and effective solution for mixed bird sound separation.
• Its high performance and efficiency make it suitable for real-time applications.
• Valuable for individual bird distinction, studying bird interactions, and automated species identification on edge devices.