Fully quantized spiking neural network for noise and echo reduction

Abstract

This thesis presents the development of a quantized spiking neural network (SNN) designed for speech enhancement, noise suppression, and echo cancellation in low-power embedded devices. With the increasing reliance on devices such as smartphones, hearing aids, and teleconferencing systems, maintaining high-quality audio is challenged by background noise and acoustic echoes. Traditional noise reduction algorithms often perform inadequately under the computational constraints typical of embedded platforms.

Spiking neural networks, leveraging event-driven computation, offer promising advantages in energy efficiency and processing demands, making them well-suited for embedded applications. However, real-time speech enhancement using SNNs remains a significant challenge. Recent advances in surrogate gradient methods, network quantization, and pruning techniques have improved the applicability of SNNs for this purpose.

This work proposes a quantized spiking neural network architecture optimized using quantization-aware training (QAT) to address the limitations in processing power and memory inherent in embedded systems. The proposed model demonstrates competitive performance in speech enhancement and echo cancellation tasks, achieving comparable results to the Deep Plastic Spiking Neural Network (DPSNN) while significantly reducing power consumption and memory footprint.

The findings confirm that with appropriate design and optimization, SNNs are capable of real-time speech enhancement, offering a compelling solution for applications in hearing aids and mobile communication systems. This research contributes to the advancement of brain-inspired computing for embedded AI, highlighting the potential of quantized spiking neural networks in speech processing applications.