An adaptive transmission line cochlear model based front-end for replay attack detection

Highlights

• We propose an adaptive transmission line cochlear model for use in speech front-ends.

• The proposed adaptive elements of the cochlear model lead to improved frequency selectivity and dynamic range compression.

• The model helps capture low amplitude channel characteristics which aid in replay detection.

Abstract

The cochlea is a remarkable spectrum analyser with desirable properties including sharp frequency tuning and level-dependent compression and the potential advantages of incorporating these characteristics in a speech processing front-end are investigated. This paper develops a framework for an active transmission line cochlear model employing adaptive notch and resonant filters. The proposed model reproduces the observed asymmetric auditory filter shape with a sharp high-frequency roll-off and level-dependent nonlinear dynamic range compression characteristics. Experimental analysis demonstrates that sharp frequency tuning and dynamic range compression of the proposed model lead to an enhanced spectral representation compared with other spectral analysis methods. The proposed model was employed in the front-end of replay spoofing attack detection systems, and experiments on the ASVspoof 2017 version 2.0 and ASVspoof 2019 databases demonstrate that the proposed model outperforms linear and nonlinear level-dependent parallel filter bank auditory models and classical spectro-temporal front-ends. The use of the proposed model leads to relative improvements of 45.6%, 51.9% and 60.8% over the baseline feature CQCCs of ASVspoof version 2.0 and CQCCs and LFCCs of ASVspoof2019 on evaluation datasets, respectively.

Introduction

Auditory model front ends are integrated into a vast majority of the speech processing systems and have been shown to outperform conventional speech processing techniques (Kim et al., 1999; Tchorz and Kollmeier, 1999). Multiple approaches to computational auditory modeling have been reported in the literature. For example, conventional auditory filters have been implemented using a set of overlapping parallel filter banks (Hohmann, 2002; Irino and Patterson, 2006). Alternatively, transmission line auditory models (Lyon, 1997) (Kates, 1991), a cascade of digital filters that closely mimic underlying cochlea physiology have also been developed. These transmission line models reproduce auditory responses more realistically than parallel filter bank models (Lyon, 2011b; Hemmert et al., 2004).

Sharp frequency tuning and nonlinear level dependent dynamic range compression are known to be two prominent phenomena responsible for the sensitivity and selectivity of the auditory systems over a broad intensity and frequency range (Moore, 1985; Robles and Ruggero, 2001). Measurements of mammalian cochlea demonstrate that the cochlea has remarkable frequency selectivity with a steep high-frequency roll-off (Moore, 1978). This improved frequency selectivity in turn could lead to noise robustness (Li, 2009).

The level-dependent nonlinear dynamic range compression is achieved via an active feedback mechanism that modifies the auditory response such that low amplitude input signals are boosted. This contributes to increasing the speech intelligibility (French and Steinberg, 1947), (Villchur, 1989).Auditory models that include level-dependent nonlinearity have been shown to improve the generalisability of speech enhancement systems (Baby and Verhulst, 2018) and have been successful in analysing, classifying and recognizing sounds in applications such as audio content categorization and music recommendation (Lyon, 2011b).

A number of active auditory models that include the level dependent nonlinearities have been validated by comparing response characteristics with the available experimental measurements of the cochlea (Walters, 2011), (Kates, 1993). However, their application in different speech processing systems has not been extensively investigated thus far.

In this paper, an active cochlear model that is focused on reproducing the sharp frequency tuning and level-dependent nonlinear characteristics of the cochlea in a way that closely matches the physiological observations is developed. A front-end based on this model is then developed for replay spoofing attack detection in automatic speaker verification systems. The channel and environmental acoustic distortions are the key discriminative cues used to identify the replay attack (Wu et al., 2015), (Singh and Pati, 2019). It is anticipated that the proposed model will effectively capture these discriminative cues from regions of silence, pauses or low speech amplitude. The proposed model is an extends earlier work published by the authors (Gunendradasan et al., 2019a; Gunendradasan et al., 2019b) to incorporate level-dependent non-linear dynamic range compression.