An explainability study of the constant Q cepstral coefficient spoofing countermeasure for automatic speaker verification

View on arXiv ← Back to list

Authors: Hemlata Tak, Jose Patino, Andreas Nautsch, Nicholas Evans, Massimiliano Todisco

Published: 2020-04-14 11:16:10+00:00

AI Summary

This paper investigates why constant Q cepstral coefficients (CQCCs) are effective in detecting some spoofing attacks but not others in automatic speaker verification. The study reveals that the effectiveness depends on the frequency location of spoofing artefacts and how different front-ends emphasize information at various frequencies.

Abstract

Anti-spoofing for automatic speaker verification is now a well established area of research, with three competitive challenges having been held in the last 6 years. A great deal of research effort over this time has been invested into the development of front-end representations tailored to the spoofing detection task. One such approach known as constant Q cepstral coefficients (CQCCs) have been shown to be especially effective in detecting attacks implemented with a unit selection based speech synthesis algorithm. Despite their success, they largely fail in detecting other forms of spoofing attack where more traditional front-end representations give substantially better results. Similar differences were also observed in the most recent, 2019 edition of the ASVspoof challenge series. This paper reports our attempts to help explain these observations. The explanation is shown to lie in the level of attention paid by each front-end to different sub-band components of the spectrum. Thus far, surprisingly little has been learned about what artefacts are being detected by spoofing countermeasures. Our work hence aims to shed light upon signal or spectrum level artefacts that serve to distinguish different forms of spoofing attack from genuine, bone fide speech. With a better understanding of these artefacts we will be better positioned to design more reliable countermeasures.

Key findings

The study found that different spoofing attacks have artefacts at different frequencies. CQCCs with linear scaling excel at detecting high-frequency artefacts, while geometrically-scaled CQCCs are better at detecting low-frequency artefacts. No single CQCC configuration performs well for all attacks, highlighting the importance of frequency-specific analysis and classifier fusion.

Approach

The researchers employed sub-band analysis by applying low-pass and high-pass filters to the audio signals at various frequencies. They then trained Gaussian Mixture Model (GMM) classifiers on these sub-bands for both LFCC and CQCC features to determine the frequency regions most informative for spoofing detection.

Datasets

ASVspoof 2019 logical access (LA) database

Model(s)

Gaussian Mixture Model (GMM) classifiers with LFCC and CQCC features

Author countries

France

← Previous