Search

Deep speaker embeddings have been demonstrated to outperform their generative counterparts, i-vectors, in recent speaker verification evaluations. To combine the benefits of high performance and generative interpretation, we investigate the use of deep embedding extractor and i-vector extractor in succession. To bundle the deep embedding extractor with an i-vector extractor, we adopt aggregation layers inspired by the Gaussian mixture model (GMM) to the embedding extractor networks. The inclusion of GMM-like layer allows the discriminatively trained network to be used as a provider of sufficient statistics for the i-vector extractor to extract what we call neural i-vectors. We compare the deep embeddings to the proposed neural i-vectors on the Speakers in the Wild (SITW) and the Speaker Recognition Evaluation (SRE) 2018 and 2019 datasets. On the core-core condition of SITW, our deep embeddings obtain performance comparative to the state-of-the-art. The neural i-vectors obtain about 50% worse performance than the deep embeddings, but on the other hand outperform the previous i-vector approaches reported in the literature by a clear margin., Comment: Accepted to Odyssey 2020: The Speaker and Language Recognition Workshop. Version 2 (bugfix)

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., Accepted to the Speaker Odyssey (The Speaker and Language Recognition Workshop) 2020 conference. 8 pages

Modern automatic speaker verification relies largely on deep neural networks (DNNs) trained on mel-frequency cepstral coefficient (MFCC) features. While there are alternative feature extraction methods based on phase, prosody and long-term temporal operations, they have not been extensively studied with DNN-based methods. We aim to fill this gap by providing extensive re-assessment of 14 feature extractors on VoxCeleb and SITW datasets. Our findings reveal that features equipped with techniques such as spectral centroids, group delay function, and integrated noise suppression provide promising alternatives to MFCCs for deep speaker embeddings extraction. Experimental results demonstrate up to 16.3\% (VoxCeleb) and 25.1\% (SITW) relative decrease in equal error rate (EER) to the baseline., Accepted to Interspeech 2020

Security of automatic speaker verification (ASV) systems is compromised by various spoofing attacks. While many types of non-proactive attacks (and their defenses) have been studied in the past, attacker's perspective on ASV, represents a far less explored direction. It can potentially help to identify the weakest parts of ASV systems and be used to develop attacker-aware systems. We present an overview on this emerging research area by focusing on potential threats of adversarial attacks on ASV, spoofing countermeasures, or both. We conclude the study with discussion on selected attacks and leveraging from such knowledge to improve defense mechanisms against adversarial attacks.

The I4U consortium was established to facilitate a joint entry to NIST speaker recognition evaluations (SRE). The latest edition of such joint submission was in SRE 2018, in which the I4U submission was among the best-performing systems. SRE'18 also marks the 10-year anniversary of I4U consortium into NIST SRE series of evaluation. The primary objective of the current paper is to summarize the results and lessons learned based on the twelve subsystems and their fusion submitted to SRE'18. It is also our intention to present a shared view on the advancements, progresses, and major paradigm shifts that we have witnessed as an SRE participant in the past decade from SRE'08 to SRE'18. In this regard, we have seen, among others , a paradigm shift from supervector representation to deep speaker embedding, and a switch of research challenge from channel compensation to domain adaptation.

We present the Voice Conversion Challenge 2018, designed as a follow up to the 2016 edition with the aim of providing a common framework for evaluating and comparing different state-of-the-art voice conversion (VC) systems. The objective of the challenge was to perform speaker conversion (i.e. transform the vocal identity) of a source speaker to a target speaker while maintaining linguistic information. As an update to the previous challenge, we considered both parallel and non-parallel data to form the Hub and Spoke tasks, respectively. A total of 23 teams from around the world submitted their systems, 11 of them additionally participated in the optional Spoke task. A large-scale crowdsourced perceptual evaluation was then carried out to rate the submitted converted speech in terms of naturalness and similarity to the target speaker identity. In this paper, we present a brief summary of the state-of-the-art techniques for VC, followed by a detailed explanation of the challenge tasks and the results that were obtained., Comment: Accepted for Speaker Odyssey 2018

The fundamental frequency (F0) contour of speech is a key aspect to represent speech prosody that finds use in speech and spoken language analysis such as voice conversion and speech synthesis as well as speaker and language identification. This work proposes new methods to estimate the F0 contour of speech using deep neural networks (DNNs) and recurrent neural networks (RNNs). They are trained using supervised learning with the ground truth of F0 contours. The latest prior research addresses this problem first as a frame-by-frame-classification problem followed by sequence tracking using deep neural network hidden Markov model (DNN-HMM) hybrid architecture. This study, however, tackles the problem as a regression problem instead, in order to obtain F0 contours with higher frequency resolution from clean and noisy speech. Experiments using PTDB-TUG corpus contaminated with additive noise (NOISEX-92) show the proposed method improves gross pitch error (GPE) by more than 25 % at signal-to-noise ratios (SNRs) between -10 dB and +10 dB as compared with one of the most noise-robust F0 trackers, PEFAC. Furthermore, the performance on fine pitch error (FPE) is improved by approximately 20 % against a state-of-the-art DNN-HMM-based approach.

The front-end factor analysis (FEFA), an extension of principal component analysis (PPCA) tailored to be used with Gaussian mixture models (GMMs), is currently the prevalent approach to extract compact utterance-level features (i-vectors) for automatic speaker verification (ASV) systems. Little research has been conducted comparing FEFA to the conventional PPCA applied to maximum a posteriori (MAP) adapted GMM supervectors. We study several alternative methods, including PPCA, factor analysis (FA), and two supervised approaches, supervised PPCA (SPPCA) and the recently proposed probabilistic partial least squares (PPLS), to compress MAP-adapted GMM supervectors. The resulting i-vectors are used in ASV tasks with a probabilistic linear discriminant analysis (PLDA) back-end. We experiment on two different datasets, on the telephone condition of NIST SRE 2010 and on the recent VoxCeleb corpus collected from YouTube videos containing celebrity interviews recorded in various acoustical and technical conditions. The results suggest that, in terms of ASV accuracy, the supervector compression approaches are on a par with FEFA. The supervised approaches did not result in improved performance. In comparison to FEFA, we obtained more than hundred-fold (100x) speedups in the total variability model (TVM) training using the PPCA and FA supervector compression approaches., To appear in Speaker Odyssey 2018: The Speaker and Language Recognition Workshop

Text-dependent automatic speaker verification naturally calls for the simultaneous verification of speaker identity and spoken content. These two tasks can be achieved with automatic speaker verification (ASV) and utterance verification (UV) technologies. While both have been addressed previously in the literature, a treatment of simultaneous speaker and utterance verification with a modern, standard database is so far lacking. This is despite the burgeoning demand for voice biometrics in a plethora of practical security applications. With the goal of improving overall verification performance, this paper reports different strategies for simultaneous ASV and UV in the context of short-duration, text-dependent speaker verification. Experiments performed on the recently released RedDots corpus are reported for three different ASV systems and four different UV systems. Results show that the combination of utterance verification with automatic speaker verification is (almost) universally beneficial with significant performance improvements being observed., published version, peerReviewed

Speech activity detection (SAD), the task of locating speech segments from a given recording, remains challenging under acoustically degraded conditions. In 2015, National Institute of Standards and Technology (NIST) coordinated OpenSAD bench-mark. We summarize “HAPPY” team effort to OpenSAD. SADs come in both unsupervised and supervised flavors, the latter requiring a labeled training set. Our solution fuses six base SADs (2 supervised and 4 unsupervised). The individually best SAD, in terms of detection cost function (DCF), is supervised and uses adaptive segmentation with i-vectors to represent the segments. Fusion of the six base SADs yields a relative decrease of 9.3% in DCF over this SAD. Further, relative decrease of 17.4% is obtained by incorporating channel detection side information., published version, peerReviewed

Accuracy of automatic speaker recognition (ASV) systems degrades severely in the presence of background noise. In this paper, we study the use of additional side information provided by a body-conducted sensor, throat microphone. Throat microphone signal is much less affected by background noise in comparison to acoustic microphone signal. This makes throat microphones potentially useful for feature extraction or speech activity detection. This paper, firstly, proposes a new prototype system for simultaneous data-acquisition of acoustic and throat microphone signals. Secondly, we study the use of this additional information for both speech activity detection, feature extraction and fusion of the acoustic and throat microphone signals. We collect a pilot database consisting of 38 subjects including both clean and noisy sessions. We carry out speaker verification experiments using Gaussian mixture model with universal background model (GMM-UBM) and i-vector based system. We have achieved considerable improvement in recognition accuracy even in highly degraded conditions., published version, peerReviewed

This study focuses in the impact of age-related intentional voice modification, or age disguise, on the performance of automatic speaker verification (ASV) systems. The data collected for this study includes 60 native Finnish speakers (29 males, 31 females) with age range between 18 and 73 years. The corpus consist of two sessions of read speech per speaker. Our experiments demonstrate vulnerability of modern ASV systems when a person attempts to conceal his or her identity, by modifying the voice to sound like an old or young person. For our i-vector PLDA system, the increase in equal error rate (EER), in the case of male speakers, was 7-fold for the attempt of old voice and 11-fold for young voice. Similar degradation in performance is observed for female speakers with a 5-fold increase in EER for old voice disguise and a 6-fold increase for young voice disguise. We further analyze the factors affecting the performance of ASV systems for the studied speech data. In our experiments, male speakers were found more successful in disguising their voices. The effect on fundamental frequency (F0) was also studied. The mean F0 distributions showed a shift towards higher frequencies when speakers attempted a young voice, which relates to the perception that younger speakers’ F0 values tend to be higher than for older speakers., published version, peerReviewed

It is well known that automatic speaker verification (ASV) systems can be vulnerable to spoofing. The community has responded to the threat by developing dedicated countermeasures aimed at detecting spoofing attacks. Progress in this area has accelerated over recent years, partly as a result of the first standard evaluation, ASVspoof 2015, which focused on spoofing detection in isolation from ASV. This paper investigates the integration of state-of-the-art spoofing countermeasures in combination with ASV. Two general strategies to countermeasure integration are reported: cascaded and parallel. The paper reports the first comparative evaluation of each approach performed with the ASVspoof 2015 corpus. Results indicate that, even in the case of varying spoofing attack algorithms, ASV performance remains robust when protected with a diverse set of integrated countermeasures., published version, peerReviewed

The performance of biometric systems based on automatic speaker recognition technology is severely degraded due to spoofing attacks with synthetic speech generated using different 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 study include previously investigated features as well as some other potentially useful features for characterizing real and synthetic 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 speech., final draft, peerReviewed

Any biometric recognizer is vulnerable to direct spoofing attacks and automatic speaker verification (ASV) is no exception; replay, synthesis and conversion attacks all provoke false acceptances unless countermeasures are used. We focus on voice conversion (VC) attacks. Most existing countermeasures use full knowledge of a particular VC system to detect spoofing. We study a potentially more universal approach involving generative modeling perspective. Specifically, we adopt standard ivector representation and probabilistic linear discriminant analysis (PLDA) back-end for joint operation of spoofing attack detector and ASV system. As a proof of concept, we study a vocoder-mismatched ASV and VC attack detection approach on the NIST 2006 speaker recognition evaluation corpus. We report stand-alone accuracy of both the ASV and countermeasure systems as well as their combination using score fusion and joint approach. The method holds promise. Index Terms: speaker recognition, spoofing, voice conversion attack, i-vector, joint verification and anti-spoofing

Voice conversion, a technique to change one’s voice to sound like that of another, poses a threat to even high performance speaker verification system. Vulnerability of text-independent speaker verification systems under spoofing attack, using statistical voice conversion technique, was evaluated and confirmed in our previous work. In this paper, we further extend the study to text-dependent speaker verification systems. In particular, we compare both joint density Gaussian mixture model (JD-GMM) and unit-selection (US) spoofing methods and, for the first time, the performances of text-independent and text-dependent speaker verification systems in a single study. We conduct the experiments using RSR2015 database which is recorded using multiple mobile devices. The experimental results indicate that text-dependent speaker verification system tolerates spoofing attacks better than the text-independent counterpart. Index Terms: Speaker verification, text-dependent, textindependent, voice conversion, spoofing attack, security

Although temporal information of speech has been shown to play an important role in perception, most of the voice conversion approaches assume the speech frames are independent of each other, thereby ignoring the temporal information. In this study, we improve conventional unit selection approach by using exemplars which span multiple frames as base units, and also take temporal information constraint into voice conversion by using overlapping frames to generate speech parameters. This approach thus provides more stable concatenation cost and avoids discontinuity problem in conventional unit selection approach. The proposed method also keeps away from the over-smoothing problem in the mainstream joint density Gaussian mixture model (JD-GMM) based conversion method by directly using target speaker’s training data for synthesizing the converted speech. Both objective and subjective evaluations indicate that our proposed method outperforms JD-GMM and conventional unit selection methods. Index Terms: Voice conversion, unit selection, multi-frame exemplar, temporal information

Accuracy of speaker verification is high under controlled conditions but falls off rapidly in the presence of interfering sounds. This is because spectral features, such as Mel-frequency cepstral coefficients (MFCCs), are sensitive to additive noise. MFCCs are a particular realization of warped-frequency representation with low-frequency focus. But there are several alternative, potentially more robust, warped-frequency representations. We provide an experimental comparison of five warped-frequency features. They use exactly the same frequency warping function, the same number of coefficients and postprocessing, but differ in their internal computations. The compared variants are (1) conventional MFCCs from discrete Fourier transform (DFT), followed by Mel-scaled filterbank, (2) MFCCs via direct warping of DFT, followed by linear-scale filterbank, (3) warped linear prediction features, (4) perceptual minimum variance distortionless features and (5) recently proposed sparse Mel-scale histogram features. Experiments carried out on a subset of the SRE 10 corpus using a scaled-down i-vector system indicate that direct DFT warping outperforms conventional MFCCs in most of the cases. Index Terms: speaker recognition, noise, frequency warping

This paper describes a novel approach to improve monoauralspeaker identification where two speakers are present in asingle-microphone recording. The goal is to identify both ofthe underlying speakers in the given mixture. The proposedapproach is composed of a double-talk detector (DTD) as a preprocessorand speaker identification back-end. We demonstratethat including the double-talk detector improves the speakeridentification accuracy. Experiments on GRID corpus show thatincluding the DTD improves average recognition accuracy from96.53% to 97.43%.

Usually the mel-frequency cepstral coefficients (MFCCs) are derived via Hamming windowed DFT spectrum. In this paper, we advocate to use a so-called multitaper method instead. Multitaper methods form a spectrum estimate using multiple window functions and frequency-domain averaging. Multitapers provide a robust spectrum estimate but have not received much attention in speech processing. Our speaker recognition experiment on NIST 2002 yields equal error rates (EERs) of 9.66 % (clean data) and 16.41 % (-10 dB SNR) for the conventional Hamming method and 8.13 % (clean data) and 14.63 % (-10 dB SNR) using multitapers. Multitapering is a simple and robust alternative to the Hamming window method.

Cohort normalization is a method for normalizing the scores in speaker verification in order to reduce undesirable variation arising from acoustically mismatched conditions. A particular form of cohort normalization, unconstrained cohort normalization (UCN) is addressed in this study. The UCN method has been shown to give excellent results but its major drawback is the huge computational load arising from the search of the cohort speakers. In this paper, we propose a fast cohort search algorithm, that quantizes the test vector sequence and uses the quantized data for both impostor and claimant scoring. Results on the NIST-1999 corpus show a speed-up factor of 23:1 compared to full search. Furthermore, the equal error rates are decreased from those of the full search.

In speaker identification, most of the computation originates from distance or likelihood computations between the feature vectors of the unknown speaker and the models in the database. The identification time depends on the number of feature vectors, their dimensionality, the complexity of the speaker models and the number of speakers. In this paper, we focus on optimizing vector quantization (VQ) based speaker identification. We reduce the number of test vectors by pre-quantizing the test sequence prior to matching, and the number of speakers by pruning out unlikely speakers during the identification process. The best variants are then generalized to Gaussian mixture model (GMM) based modeling also. We obtain a speed-up factor of 16:1 with VQ-based system, and 34:1 with GMM-based system with a minor degradation in the identification error rate.