Towards End-to-End Acoustic Localization Using Deep Learning: From Audio Signals to Source Position Coordinates.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Accurate indoor acoustic source localization is crucial for various applications.
• Existing methods often rely on handcrafted features or complex algorithms.
• Limited labeled data poses a challenge for training deep learning models.

Purpose of the Study:

• To develop a novel Convolutional Neural Network (CNN) for direct 3D acoustic source localization.
• To address the challenge of limited training data using a semi-synthetic and real data fine-tuning strategy.
• To evaluate the proposed CNN's performance against established and recent localization techniques.

Main Methods:

• A CNN model designed to directly process raw audio signals for 3D position estimation.
• A two-step training strategy involving semi-synthetic data generation and real-data fine-tuning.
• Simulation of signal propagation delays and distortions for semi-synthetic data creation.

Main Results:

• The proposed CNN approach significantly improves localization accuracy compared to SRP-PHAT and CRNN methods.
• The model demonstrates robust performance irrespective of speaker gender or signal window size.
• Effective localization achieved even with limited real-world training data.

Conclusions:

• The novel CNN-based approach offers a superior method for indoor acoustic source localization.
• The proposed training strategy effectively overcomes data scarcity issues.
• This method provides a promising direction for enhancing real-time acoustic sensing applications.