Spectral Root Features for Replay Spoof Detection in Voice Assistants

Abstract

In this paper, authors propose spectral root cepstral coefficients (SRCC) feature set to develop the effective countermeasure system for replay attacks on voice assistants (VAs). Experiments are performed on ReMASC dataset, which is specifically designed for the replay attack detection task. Logarithm operation in MFCC extraction is replaced by power-law nonlinearity (i.e. \((\cdot )^\gamma \)) to derive SRCC feature set. The proper choice of the \(\gamma \) helps to capture the system information of the speech signal, with a minimum number of cepstral coefficients. We investigated two approaches for proper choice of \(\gamma \)-value, in particular, by estimating the energy concentration in cepstral coefficients and by visualizing the spectrogram w.r.t. \(\gamma \)-value. This system representation of the speech signal, is the discriminative cue for the replay spoof speech detection (SSD) task as replay speech signal consists of additional transmission channel effects convolved with the genuine signal. The performance of the proposed feature set is validated using Gaussian Mixture Model (GMM), and Light Convolutional Neural Network (LCNN). Our primary system shows relative improvement of 47.49% over the baseline system (Constant-Q Cepstral Coefficients (CQCC)-GMM) on the evaluation set. The EER is further reduced to 11.84% on evaluation set by classifier-level fusion.