Low-power Spiking Neural Network audio source localisation using a Hilbert Transform audio event encoding scheme.

Area of Science:

• Signal Processing
• Computational Neuroscience
• Acoustics

Background:

• Sound source localization is crucial for noise reduction in consumer electronics.
• Traditional beamforming methods require dense filters and are computationally intensive, limiting their use in low-power devices.
• Spiking neural networks (SNNs) offer potential for efficient audio processing but require specialized algorithms.

Purpose of the Study:

• To develop an efficient sound source localization method suitable for ultra-low-power spiking neural networks (SNNs).
• To reduce the computational demands of traditional beamforming techniques.
• To enable high-accuracy audio source separation in power-constrained Internet of Things (IoT) devices.

Main Methods:

• Utilized a Hilbert transform to process wideband audio without dense band-pass filters.
• Developed an event-based encoding method to capture the phase of the complex analytic signal.
• Implemented the localization method on arbitrary microphone arrays for efficient SNN processing.

Main Results:

• Achieved high accuracy in sound source localization comparable to traditional super-resolution beamforming methods.
• Demonstrated significantly lower power consumption when deployed on low-power SNN inference hardware.
• Validated the effectiveness of co-designing signal processing with SNNs for improved power efficiency.

Conclusions:

• The proposed Hilbert-transform-based beamforming method offers a computationally efficient and accurate solution for sound source localization.
• This approach is well-suited for ultra-low-power SNNs, enabling advanced audio processing in edge devices.
• The method also shows potential for enhancing the efficiency of traditional digital signal processing.