Configuration-Invariant Sound Localization Technique Using Azimuth-Frequency Representation and Convolutional Neural...

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Deep neural networks (DNNs) are advancing speech processing and sound localization.
• Current DNN models for sound localization require fixed input sizes, necessitating retraining when microphone configurations change.
• This limitation hinders the adaptability of DNN-based sound localization systems.

Discussion:

• This paper proposes a configuration-invariant sound localization technique utilizing azimuth-frequency representation and convolutional neural networks (CNNs).
• The proposed CNN model accepts azimuth-frequency representations, differing from conventional time-frequency features.
• This approach addresses the challenge of adapting DNN models to diverse microphone configurations without retraining.

Key Insights:

• The novel CNN model demonstrates superior sound localization performance compared to traditional methods like Steered Response Power Phase Transform (SRP-PHAT) and Multiple Signal Classification (MUSIC).
• Evaluations confirmed the model's effectiveness across different microphone configurations than those used during training.
• The azimuth-frequency representation proves to be a robust input feature for configuration-invariant sound localization.

Outlook:

• Further research could explore real-world acoustic environments and more complex sound scenarios.
• This technique holds potential for applications in robotics, augmented reality, and intelligent audio systems.
• Investigating the model's performance with multiple sound sources could expand its utility.