Hybrid Deep Filtering with LogNNet: Complex Spectral Filters for Robust Speech Denoising

This paper introduces Hybrid Deep Filtering with LogNNet for robust speech denoising, evaluated on synthetic white noise spanning SNR ≈ 0.3–20 dB and on real-world NOIZEUS noises (≈ 5 dB SNR). Across all levels, we observe a substantial reduction of residual noise quantified by a new integral spectral metric—the Spectral Filtration Coefficient (SFC). We further show a monotonic relation between SFC and the Pearson spectral similarity of noise to clean speech, yielding a simple rule-of-thumb predictor of denoising difficulty at fixed SNR. In a hybrid pipeline, pairing LogNNet with classical filters (Kalman, Wiener, Savitzky–Golay, Butterworth and Wavelet) consistently outperforms standalone baselines, reaching up to +60% PESQ on real-world station noise and up to +150% under light white noise; HVG features provide additional gains in some settings. Finally, we demonstrate training-time spectral shaping—mixing target speech with white noise—to obtain application-specific complex spectral filters whose passbands align with the target spectrum, supporting lightweight deployment on edge/IoT devices.