Eliminating quantization errors in classification-based sound source localization.

Area of Science:

• Acoustics and Signal Processing
• Machine Learning for Audio Analysis
• Computational Auditory Scene Analysis

Background:

• Sound Source Localization (SSL) estimates the Direction of Arrival (DOA) of sound sources.
• While regression offers precision for continuous DOA, classification is more robust but suffers from quantization error.
• Existing classification methods for DOA estimation do not fully exploit inherent inter-class correlations.

Purpose of the Study:

• To eliminate quantization error in training targets for DOA estimation using Unbiased Label Distribution (ULD).
• To overcome quantization error during the decoding stage with Weighted Adjacent Decoding (WAD).
• To enhance the performance of classification-based DOA estimation by addressing its inherent limitations.

Main Methods:

• Proposed Unbiased Label Distribution (ULD) to create unbiased training targets, mitigating quantization error.
• Introduced Weighted Adjacent Decoding (WAD) to address quantization error during the decoding phase.
• Developed two novel loss functions, Negative Log Absolute Error (NLAE) and Mean Squared Error without activation (MSE(wo)), for soft labels.

Main Results:

• The proposed ULD and WAD methods significantly reduce quantization error in DOA estimation.
• The approach surpasses traditional classification limits, demonstrating superior robustness and precision.
• Achieved state-of-the-art performance in sound source localization tasks.

Conclusions:

• ULD and WAD effectively address the quantization error inherent in classification-based DOA estimation.
• The developed methods offer a more precise and robust approach to sound source localization.
• The research provides a significant advancement in DOA estimation techniques, with code available for reproducibility.