HybrA: Learnable, interpretable, and stable filterbanks for feature extraction

When designing machine learning models, especially for acoustics applications, there is an ever-repeating question: how should feature extraction be done? The two most common approaches are to use a hand-crafted fixed filterbank or to process the audio signal directly in an end-to-end manner. Each has its own pros and cons, e.g., the first is not adaptive, and the second is poorly interpretable. We present a hybrid approach that combines the strengths of both. This approach involves convolving each filter of a fixed filterbank with a short learned kernel. In this way, we obtain a new hybrid filterbank that keeps (some) interpretability while being able to adapt to the problem at hand. To ensure that no information is lost during this initial processing stage, we additionally incorporate a regularizer that maintains stability. We present a concrete example of constructing a hybrid filterbank using an audlet transform—an invertible, perception-related time-frequency transform—and demonstrate its effectiveness in a denoising task.