Ensemble Models for Spoofing Detection in Automatic Speaker Verification

Abstract

Spectrograms -time-frequency representations of audio signals -have found widespread use in neural network-based spoofing detection. While deep models are trained on the fullband spectrum of the signal, we argue that not all frequency bands are useful for these tasks. In this paper, we systematically investigate the impact of different subbands and their importance on replay spoofing detection on two benchmark datasets: ASVspoof 2017 v2.0 and ASVspoof 2019 PA. We propose a joint subband modelling framework that employs n different sub-networks to learn subband specific features. These are later combined and passed to a classifier and the whole network weights are updated during training. Our findings on the ASVspoof 2017 dataset suggest that the most discriminative information appears to be in the first and the last 1 kHz frequency bands, and the joint model trained on these two subbands shows the best performance outperforming the baselines by a large margin. However, these findings do not generalise on the ASVspoof 2019 PA dataset. This suggests that the datasets available for training these models do not reflect real world replay conditions suggesting a need for careful design of datasets for training replay spoofing countermeasures.

Deep generative variational autoencoding for replay spoof detection in automatic speaker verification

“…While training VAEs with an auxiliary classifier on the µ z input, we use 32 neuron units on the FC layer. We do not use the entire training and development audio files for training and model validation on the ASVspoof [76] that showed good generalisation on the ASVspoof 2019 test dataset during the recent ASVspoof 2019 evaluations. Note, however, that all the evaluation portion results are reported on the standard ASVspoof protocols.…”

“…Furthermore, attentionbased models have been studied in [61,62] during the ASVspoof 2019 challenge. It is also worth noting that the best performing models on the ASVspoof challanges used fusion approaches, either at the classifier output or the feature level [57,76,15], indicating the challenges in designing a single countermeasure capable of capturing all the variabilities that may appear in wild test conditions in a presentation attack. Please refer to Table 1 for details.…”

“…We consider both CQCC [66] and log-power spectrogram features. We apply a pre-processing step on the raw-audio waveforms to trim silence/noise before and after the utterance in the training, development and test sets, following recommendations in [45] and [76]. Following [84], we extract log energy plus 19dimensional static coefficients augmented with deltas and double-deltas, yielding 60-dimensional feature vectors.…”

“…We would like to stress that we do not use the original training and development protocols for model training and validation. Instead, we use custom, but publicly released protocols available at https://github.com/BhusanChettri/ASVspoof2019 from our prior work[76] that helped to improve generalisation during the ASVspoof 2019 challenge. However, during testing, we report test results on the standard development and evaluation protocols.…”

Discriminative features based on modified log magnitude spectrum for playback speech detection

“…In which, the corpus has three subset, train, development, and evaluation set. According to ASVspoof 2019 challenge rule, tandem detection cost function (t-DCF) [56] and EER are used as the primary and secondary metric, respectively, which is the same as the previous works [57][58][59][60][61][62][63][64]. Table 10 gives the experimental results on ASVspoof 2019 physical access development set using dynamic features of CMOC and CVOC.…”