Wavelet Transform Enhanced Signal Processing for Low-Latency Audio Recognition in Edge Devices
Abstract
Low-latency audio recognition in edge devices is critical for real-time applications such as voice assistants, surveillance, and wearable tech. Wavelet Transform has emerged as a powerful tool in signal processing due to its ability to analyze non-stationary signals at multiple resolutions. However, existing methods often suffer from high computational costs, limited noise resilience, and suboptimal performance under constrained edge environments, making them less suitable for real-time deployment. To address these challenges, this paper introduces a novel framework named WESP-LAR (Wavelet Enhanced Signal Processing for Low-latency Audio Recognition), which leverages Discrete Wavelet Transform (DWT) for efficient time-frequency feature extraction and integrates lightweight neural models for classification. WESP-LAR is optimized for low-memory usage and real-time execution on edge processors. The proposed method enhances audio feature representation while significantly reducing latency and computational overhead, making it ideal for embedded AI scenarios. Experimental results demonstrate that WESP-LAR achieves superior accuracy and robustness across diverse audio datasets, outperforming traditional FFT-based and raw waveform approaches.