Figure 1: LA access and PA access adopted for ASVspoof 2019 Figure 2: Partitions and protocols of ASVspoof 2019 database. Top part shows divisions of training, development, and evaluation sets common to logical access and physical access conditions and numbers of speakers included in each set. Middle part shows database partitions for logical access condition, and bottom part shows those for physical access condition. Table 1: Summary of LA spoofing systems. * indicates neural networks. For abbreviations in this table, please refer to Section 3. Note that A04 and A16 use same waveform concatenation TTS algorithm, and A06 and A19 use same VC algorithm. Figure 3: Visualization of bona fide and spoofed speech data of ASVspoof 2019 LA subset. Black circle denotes range of bona fide data within mean +3 standard deviation. Figure 4: Agglomerative clustering of LA attacks using same input data as Figure 3. Figure 5: Illustration of AS Vspoof 2019 physical access (PA) scenario. Replay attacks are simulated within an acoustic environment/room of dimensions x x y (7.5m in the example above) with controllable reverberation. Recordings of bona fide presentations are acquired from distance D, from the talker. Bona fide or replay presentations are made at distance D, from ASV microphone. Table 2: Environment is defined as triplet (S,R,D;), each element of which takes one value in set (a,b,c) as a categorical value. Table 3: Replay attack is defined as duple (Dz,Q), each element of which takes one value in set (A,B,C) as a categorical value. Table 4: Definition of replay device quality (Q). OB refers to occupied band- width, minF is lower bound of the OB, LNLR is linear-to-non-linear power ratio. 4.2. Replay devices Table 5: List of real devices from which measurements were taken for sim- ulation of replay attack presentation. Q indicates device quality (B high, C low). Device code signifies device type: bluetooth (BT); headphone (H); mo- bile smartphone (M); larger consumer and professional loudspeaker (LS); tablet (T); laptop (LT). Level indicates volume (high, low) used during device charac- terisation. Right-most column indicates whether measured device characteris- tics were used for the simulation of utterances in training and development set (known devices) or evalutaion set (unknown devices). Figure 6: Illustration of set of higher harmonic frequency responses (HHFRs) for arbitrary smart-tablet device estimated using synchronized-swept-sine approach to nonlinear system identification based on nonlinear convolution. H is linear component, while Hz — Hs are higher order non-linear components. Blue/shaded region is occupied bandwidth, difference in frequency between points where integrated power crosses 0.5% and 99.5% of total power in spectrum. Figure 8: Illustration of PA simulation process based on impulse response (IR) modeling approach. Simulations take into account size of acoustic environ- ment S and level of reverberation R. Bona fide access attempts are made at distance D, from ASV microphone, whereas surreptitious recordings are made at distance D, from talker before being presented to ASV microphone, also at distance D,. Effects of digital-to-analogue conversion, signal amplification and replay (using loudspeaker) are all modelled, and represented with single device quality indicator Q. ‘igure 7: Characteristics measured from 40 different loudspeaker devices listed in Table 5. The top plot shows the operational bandwidth (OB). The middle plot hows the lower bound of the OB (minF). The bottom plot shows the linear-to-non-linear power ratio (LNLR) in the range of the OB. Change label of lower plot to -NLR. Figure 9: As for Figure 3, except for PA data. Bona fide utterances and re- played versions according to 9 different replay configurations are completely overlapping. Mini-clusters correspond to speaker-utterance combinations. Table 6: ASV performance in terms of EER (%) on LA subset of ASVSpoof 2019 dataset for baseline and different attack conditions for development and evaluation set. Note that A16 used same TTS algorithm as A04, and A19 used same VC algorithm as A06. expectation-maximization (EM) algorithm. Finally, scores are the log-likelihood ratio between the two hypotheses, namely that a given trial is either bona fide or spoofed speech. Base- line CMs are trained separately for LA and PA scenarios using designated CM training data. Figure 10: ASV score distributions for target bona fide, non-target bona fide, and spoofed data from A1l6. Vertical black line denote classification thresh- old between spoofed and target bona fide data. Dot and square shaded areas correspond to false reject errors and false accept errors, respectively. Table 7: Performance of integrated system in terms of min-tDCF and of stan- dalone countermeasures in terms of EER (%) on development set (LA subset) of ASVSpoof 2019 dataset. Results are shown for two baselines, B1 (CQCC- GMM) and B2 (LFCC-GMM), separately, combined with fixed ASV system based on x-vector. Last row describes results for “pooled condition” when tri- als from all the attacks are considered for evaluation. Figure 12: Summary of LA subset (evaluation section) results in terms of t-DCF (left) and EER (right). First row shows results for three categories of synthetic speech: (i) TTS, (ii) VC (TTS), and (iii) VC (Human). Next row shows results for four types of acoustic models: (i) neural-network-based pipeline TTS, (ii) neural- network-based end-to-end TTS, (iii) neural-network-based VC, and (iv) statistical-model-based VC. Last row shows results for different waveform generation methods: (i) neural waveform models, (ii) classic speech vocoders, (iii) waveform concatenation, (iv) spectral filtering, (v) waveform filtering, and (vi) others. Figure 13: An illustration of baseline results for the PA scenario of the ASVspoof 2019 database. Results illustrated for individual replay configurations (pooled acoustic environments) and for: (top panel) standalone ASV results in terms of EER (%) with target and zero-effort impostor trials (black bars) and target and replay spoofing trials (gray bars); (middle panel) standalone replay spoofing in terms of EER (%) for baselines B1 and B2; (bottom panel) combined ASV and CM results illustrated in terms of the min-tDCF. Figure 14: As for Figure 13 except for results in terms of individual acoustic environments (pooled replay configurations). Figure 15: DET curves based on human assessment of similarity to target speakers (left) and speech quality (right)
