Knowledge-Based Wavelet Filters Prominently Detect Spoofed Speech

Abstract

Although voice biometrics has been at the forefront of speech technologies, spoofing attacks have been one of the main issues responsible for avoiding its practical usage in commercial applications. Consequently, this article presents our proposed approach for building Knowledge-based wavelet filters, particularly dedicated to distinguish between genuine and spoofed speech. The main contribution of our strategy, particularly dedicated to identify replay attacks, is the acquisition of knowledge through the so called knowledge coefficients. Our results, obtained upon performing a set of experiments, indicate that the designed wavelet filters successfully classified hundreds of speech signals from our dataset, stimulating further research in this direction.