Bridging DSP and DL for Speech Enhancement

The pursuit of truly adaptive and interpretable speech enhancement models has long been a critical challenge. Traditional Digital Signal Processing (DSP) methods offer interpretability but struggle with dynamic, non-stationary noise. Conversely, deep learning excels at adaptation but often operates as a 'black box'. A new approach, TVF (Time-Varying Filtering), emerges to bridge this divide.

Neural Coefficients for Adaptive IIR Filters

TVF introduces a novel architecture that leverages a lightweight neural network to predict the coefficients for a cascade of 35-band Infinite Impulse Response (IIR) filters. This differentiable design allows the filtering process to adapt dynamically in real-time to changing acoustic environments, a significant leap from static filtering techniques. The resulting Time-Varying Filtering speech enhancement model boasts approximately 1 million parameters, striking a balance between performance and computational efficiency.

Interpretable Spectral Control

Unlike end-to-end deep learning solutions, TVF maintains complete interpretability. The spectral modifications are explicit and directly adjustable through the predicted filter coefficients. This transparency is crucial for debugging, fine-tuning, and gaining deeper insights into the speech enhancement process. The researchers demonstrated TVF's efficacy on a speech denoising task, showing its ability to adapt to changing noise conditions effectively when compared to static DDSP and fully deep-learning-based methods.