The performance of biometric systems based on automatic speaker recognition technology is severely degraded due to spoofing attacks with synthetic speech generated using diff erent voice conversion (VC) and speech synthesis (SS) techniques. Various countermeasures are proposed to detect this type of attack, and in this context, choosing an appropriate feature extraction technique for capturing relevant information from speech is an important issue. This paper presents a concise experimental review of different features for synthetic speech detection task. A wide variety of features considered in this stud y include previously investigated features as well as some other potentially useful features for characterizing real and sy nthetic speech. The experiments are conducted on recently released ASVspoof 2015 corpus containing speech data from a large number of VC and SS technique. Comparative results using two different classifiers indicate that features representing spectral information in high-frequency region, dynamic information of speech, and detailed information related to subband characteristics are considerably more useful in detecting synthetic sp eech. Index Terms: anti-spoofing, ASVspoof 2015, feature extraction, countermeasures

A Comparison of Features for Synthetic Speech Detection

Text-independent speaker verification against short utterances is still challenging despite of recent advances in the field of speaker recognition with i-vector framework. In general, to get a robust i-vector representation, a satisfying amount of data is needed in the MAP adaptation step, which is hard to meet under short duration constraint. To overcome this, we present an endto-end system which directly learns a mapping from speech features to a compact fixed length speaker discriminative embedding where the Euclidean distance is employed for measuring similarity within trials. To learn the feature mapping, a modified Inception Net with residual block is proposed to optimize the triplet loss function. The input of our end-to-end system is a fixed length spectrogram converted from an arbitrary length utterance. Experiments show that our system consistently outperforms a conventional i-vector system on short duration speaker verification tasks. To test the limit under various duration conditions, we also demonstrate how our end-to-end system behaves with different duration from 2s-4s.

End-to-End Text-Independent Speaker Verification with Triplet Loss on Short Utterances.

This paper presents our contribution to the ASVspoof 2017 Challenge. It addresses a replay spoofing attack against a speaker recognition system by detecting that the analysed signal has passed through multiple analogue-to-digital (AD) conversions. Specifically, we show that most of the cues that enable to detect the replay attacks can be found in the high-frequency band of the replayed recordings. The described anti-spoofing countermeasures are based on (1) modelling the subband spectrum and (2) using the proposed features derived from the linear prediction (LP) analysis. The results of the investigated methods show a significant improvement in comparison to the baseline system of the ASVspoof 2017 Challenge. A relative equal error rate (EER) reduction by 70% was achieved for the development set and a reduction by 30% was obtained for the evaluation set.

Audio Replay Attack Detection Using High-Frequency Features.

Background
Machine learning (ML) provides novel opportunities for human behavior research and clinical translation, yet its application can have noted pitfalls (Bone et al., 2015). In this work, we fastidiously utilize ML to derive autism spectrum disorder (ASD) instrument algorithms in an attempt to improve upon widely used ASD screening and diagnostic tools.

Methods
The data consisted of Autism Diagnostic Interview-Revised (ADI-R) and Social Responsiveness Scale (SRS) scores for 1,264 verbal individuals with ASD and 462 verbal individuals with non-ASD developmental or psychiatric disorders, split at age 10. Algorithms were created via a robust ML classifier, support vector machine, while targeting best-estimate clinical diagnosis of ASD versus non-ASD. Parameter settings were tuned in multiple levels of cross-validation.

Results
The created algorithms were more effective (higher performing) than the current algorithms, were tunable (sensitivity and specificity can be differentially weighted), and were more efficient (achieving near-peak performance with five or fewer codes). Results from ML-based fusion of ADI-R and SRS are reported. We present a screener algorithm for below (above) age 10 that reached 89.2% (86.7%) sensitivity and 59.0% (53.4%) specificity with only five behavioral codes.

Conclusions
ML is useful for creating robust, customizable instrument algorithms. In a unique dataset comprised of controls with other difficulties, our findings highlight the limitations of current caregiver-report instruments and indicate possible avenues for improving ASD screening and diagnostic tools.

Use of machine learning to improve autism screening and diagnostic instruments: effectiveness, efficiency, and multi-instrument fusion

Speaker recognition systems trained on long duration utterances are known to perform significantly worse when short test segments are encountered. To address this mismatch, we analyze the effect of duration variability on phoneme distributions of speech utterances and i-vector length. We demonstrate that, as utterance duration is decreased, number of detected unique phonemes and i-vector length approaches zero in a logarithmic and non-linear fashion, respectively. Assuming duration variability as an additive noise in the i-vector space, we propose three different strategies for its compensation: i) multi-duration training in Probabilistic Linear Discriminant Analysis (PLDA) model, ii) score calibration using log duration as a Quality Measure Function (QMF), and iii) multi-duration PLDA training with synthesized short duration i-vectors. Experiments are designed based on the 2012 National Institute of Standards and Technology (NIST) Speaker Recognition Evaluation (SRE) protocol with varying test utterance duration. Experimental results demonstrate the effectiveness of the proposed schemes on short duration test conditions, especially with the QMF calibration approach.

/pdf/duration-mismatch-compensation-for-i-vector-based-speaker-xkb5igafum.pdf

Duration mismatch compensation for i-vector based speaker recognition systems

In this study, we present systems submitted by the Center for Robust Speech Systems (CRSS) from UTDallas to NIST SRE 2018 (SRE18). Three alternative front-end speaker embedding frameworks are investigated, that includes: (i) i-vector, (ii) x-vector, (iii) and a modified triplet speaker embedding system (t-vector). Similar to the previous SRE, language mismatch between training and enrollment/test data, the so-called domain mismatch, remains as a major challenge in this evaluation. In addition, SRE18 also introduces a small portion of audio from an unstructured video corpus in which speaker detection/diarization is supposedly needed to be effectively integrated into speaker recognition for system robustness. In our system development, we focused on: (i) building novel deep neural network based speaker discriminative embedding systems as utterance level feature representations, (ii) exploring alternative dimension reduction methods, back-end classifiers, score normalization techniques which can incorporate unlabeled in-domain data for domain adaptation, (iii) finding an improved data set configurations for the speaker embedding network, LDA/PLDA, and score calibration training (v) and finally, investigating effective score calibration and fusion strategies. The final resulting systems are shown to be both complementary and effective in achieving overall improved speaker recognition performance.

/pdf/utd-crss-systems-for-2018-nist-speaker-recognition-4q0fhu5kd4.pdf

UTD-CRSS Systems for 2018 NIST Speaker Recognition Evaluation

Speaker verification systems often degrade significantly when there is a language mismatch between training and testing data. Being able to improve cross-lingual speaker verification system using unlabeled data can greatly increase the robustness of the system and reduce human labeling costs. In this study, we introduce an unsupervised Adversarial Discriminative Domain Adaptation (ADDA) method to effectively learn an asymmetric mapping that adapts the target domain encoder to the source domain, where the target domain and source domain are speech data from different languages. ADDA, together with a popular Domain Adversarial Training (DAT) approach, are evaluated on a cross-lingual speaker verification task: the training data is in English from NIST SRE04-08, Mixer 6 and Switchboard, and the test data is in Chinese from AISHELL-I. We show that with the ADDA adaptation, Equal Error Rate (EER) of the x-vector system decreases from 9.331% to 7.645%, relatively 18.07% reduction of EER, and 6.32% reduction from DAT as well. Further data analysis of ADDA adapted speaker embedding shows that the learned speaker embeddings can perform well on speaker classification for the target domain data, and are less dependent with respect to the shift in language.

Cross-lingual Text-independent Speaker Verification Using Unsupervised Adversarial Discriminative Domain Adaptation

In this study, we investigate self-supervised representation learning for speaker verification (SV). First, we examine a simple contrastive learning approach (SimCLR) with a momentum contrastive (MoCo) learning framework, where the MoCo speaker embedding system utilizes a queue to maintain a large set of negative examples. We show that better speaker embeddings can be learned by momentum contrastive learning. Next, alternative augmentation strategies are explored to normalize extrinsic speaker variabilities of two random segments from the same speech utterance. Specifically, augmentation in the waveform largely improves the speaker representations for SV tasks. The proposed MoCo speaker embedding is further improved when a prototypical memory bank is introduced, which encourages the speaker embeddings to be closer to their assigned prototypes with an intermediate clustering step. In addition, we generalize the self-supervised framework to a semi-supervised scenario where only a small portion of the data is labeled. Comprehensive experiments on the Voxceleb dataset demonstrate that our proposed self-supervised approach achieves competitive performance compared with existing techniques, and can approach fully supervised results with partially labeled data.

Self-supervised Text-independent Speaker Verification using Prototypical Momentum Contrastive Learning

Automatically evaluating pronunciation quality of non-native speech has seen tremendous success in both research and commercial settings, with applications in L2 learning. In this paper, submitted for the INTERSPEECH 2015 Degree of Nativeness Sub-Challenge, this problem is posed under a challenging crosscorpora setting using speech data drawn from multiple speakers from a variety of language backgrounds (L1) reading different English sentences. Since the perception of non-nativeness is realized at the segmental and suprasegmental linguistic levels, we explore a number of acoustic cues at multiple time scales. We experiment with both data-driven and knowledge-inspired features that capture degree of nativeness from pauses in speech, speaking rate, rhythm/stress, and goodness of phone pronunciation. One promising finding is that highly accurate automated assessment can be attained using a small diverse set of intuitive and interpretable features. Performance is further boosted by smoothing scores across utterances from the same speaker; our best system significantly outperforms the challenge baseline.

/pdf/automated-evaluation-of-non-native-english-pronunciation-2ttqxdabre.pdf

Automated evaluation of non-native English pronunciation quality: combining knowledge- and data-driven features at multiple time scales.

Observational therapy is an important element of mental health that relies on a detailed assessment of multiple behavioral cues. Behavioral coding for research in the field is unfortunately often at session-level resolution due to the inherent cost of labeling and human subjectivity. Being able to model the interlocutors’ behavior at a fine temporal resolution and analyze the effect of such behavioral changes in the gestalt perception can help psychologists better understand the behavioral mechanism. In this paper, we propose a method to model the dynamically evolving behavior of interlocutors during couple interactions. We firstly present a static behavioral model based on the local decisions with global fusion, and investigate the impact of the frame length to provide effective global evaluations. We then propose a two-layer sequential Hidden Markov Model to capture local state transitions. We use the corpus of Couple Therapy interactions as a case study, finding that an interlocutor does not express a single behavior throughout a conversation, and there are temporal correlations between neighboring frames. We show that dynamic models can achieve up to 10% relative improvement, compared to static models. This suggests that the human behavioral interaction is a non-linear process, and the resulting latent-state labels may provide new insights to domain experts.

Wei Xia

Papers

UTD-CRSS Systems for 2018 NIST Speaker Recognition Evaluation

Cross-lingual Text-independent Speaker Verification Using Unsupervised Adversarial Discriminative Domain Adaptation

Self-supervised Text-independent Speaker Verification using Prototypical Momentum Contrastive Learning

Automated evaluation of non-native English pronunciation quality: combining knowledge- and data-driven features at multiple time scales.

A dynamic model for behavioral analysis of couple interactions using acoustic features.