Adaptive Audio Enhancement Using Recursive Dynamic Noise Tracking for Low-Latency Edge Devices

Abstract

Environmental noise can substantially affect speech quality and intelligibility in real-time audio systems, particularly when speech enhancement is implemented on computationally limited edge platforms. Conventional frequency-domain techniques, including spectral subtraction and Wiener filtering, provide computationally efficient solutions, but their effectiveness can decrease when the acoustic noise varies with time. Deep learning-based enhancement methods can achieve stronger noise suppression, although their computational, memory, and latency requirements may restrict deployment on lightweight embedded hardware. This work presents a Recursive Dynamic Noise Tracking (RDNT) framework with adaptive spectral gain estimation to provide low-latency speech enhancement under varying noise conditions. The proposed framework processes speech in the Short-Time Fourier Transform (STFT) domain and recursively updates the noise spectrum using successive noisy-speech observations. The resulting noise estimate is used to determine frequency-dependent suppression gains, while the original phase is retained during speech reconstruction. Evaluation was conducted using clean speech from the VoiceBank dataset combined with Gaussian, traffic, and crowd noise, with environmental recordings obtained from the DEMAND dataset. RDNT was compared with spectral subtraction and Wiener filtering using SNR, PESQ, STOI, SDR, SI-SDR, processing latency, and subjective Mean Opinion Score (MOS). Across the evaluated conditions, the proposed method improved the measured speech-enhancement performance while preserving intelligibility and maintaining a maximum processing latency of 0.082 s. These results demonstrate that recursive noise tracking can provide a practical compromise among noise suppression, speech preservation, and computational efficiency for real-time speech enhancement on edge platforms.