Toward learning robust contrastive embeddings for binaural sound source localization.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Deep neural networks offer accurate binaural source localization but depend heavily on training data distribution.
• Existing data-driven models map binaural cues directly to source locations, limiting generalization.

Purpose of the Study:

• To propose a parametric embedding for mapping binaural cues to a low-dimensional space for improved source localization.
• To enhance model interpretability and generalization across diverse acoustic environments.

Main Methods:

• Implemented a neural network to create a parametric embedding of binaural cues.
• Optimized the network to preserve source proximity in the embedding space, forming a manifold.
• Evaluated performance against unsupervised embeddings and feed-forward neural networks.

Main Results:

• The proposed embedding generalizes well to reverberant conditions not seen during training.
• Achieved better or equal performance compared to feed-forward neural networks, especially with limited data.
• Weakly supervised learning yielded comparable results to supervised learning, enabling simultaneous azimuth and elevation estimation.

Conclusions:

• The parametric embedding provides an interpretable and generalizable approach to binaural source localization.
• This method offers a robust alternative to direct mapping, particularly beneficial in challenging acoustic scenarios and with limited data.