The performance of a state-of-the-art speaker verification system is severely degraded when it is presented with trial recordings of short duration. In this work we propose to use deep neural networks to learn short-duration speaker embeddings. We focus on the 5s-5s condition, wherein both sides of a verification trial are 5 seconds long. In our previous work we established that learning a non-linear mapping from i-vectors to speaker labels is beneficial for speaker verification [1]. In this work we take the idea of learning a speaker classifier one step further-we apply deep neural networks directly to timefrequency speech representations. We propose two feedforward network architectures for this task. Our best model is based on a deep convolutional architecture wherein recordings are treated as images. From our experimental findings we advocate treating utterances as images or 'speaker snapshots, much like in face recognition. Our convolutional speaker embeddings perform significantly better than i-vectors when scoring is done using cosine distance, where the relative improvement is 23.5%. The proposed deep embeddings combined with cosine distance also outperform a state-of-the-art i-vector verification system by 1%, providing further empirical evidence in favor of our learned speaker features.

The duration of speech segments has traditionally been controlled in the NIST speaker recognition evaluations so that researchers working in this framework have been relieved of the responsibility of dealing with the duration variability that arises in practical applications. The fixed dimensional i-vector representation of speech utterances is ideal for working under such controlled conditions and ignoring the fact that i-vectors extracted from short utterances are less reliable than those extracted from long utterances leads to a very simple formulation of the speaker recognition problem. However a more realistic approach seems to be needed to handle duration variability properly. In this paper, we show how to quantify the uncertainty associated with the i-vector extraction process and propagate it into a PLDA classifier. We evaluated this approach using test sets derived from the NIST 2010 core and extended core conditions by randomly truncating the utterances in the female, telephone speech trials so that the durations of all enrollment and test utterances lay in the range 3-60 seconds and we found that it led to substantial improvements in accuracy. Although the likelihood ratio computation for speaker verification is more computationally expensive than in the standard i-vector/PLDA classifier, it is still quite modest as it reduces to computing the probability density functions of two full covariance Gaussians (irrespective of the number of the number of utterances used to enroll a speaker).

In this paper we study the performance of the low-variance multi-taper Mel-frequency cepstral coefficient (MFCC) and perceptual linear prediction (PLP) features in a state-ofthe-art i-vector speaker verification system. The MFCC and PLP features are usually computed from a Hamming-windowed periodogram spectrum estimate. Such a singletapered spectrum estimate has large variance, which can be reduced by averaging spectral estimates obtained using a set of different tapers, leading to a so-called multitaper spectral estimate. The multi-taper spectrum estimation method has proven to be powerful especially when the spectrum of interest has a large dynamic range or varies rapidly. Multi-taper MFCC features were also recently studied in speaker verification with promising preliminary results. In this study our primary goal is to validate those findings using an up-to-date i-vector classifier on the latest NIST 2010 SRE data. In addition, we also propose to compute robust perceptual linear prediction (PLP) features using multitapers. Furthermore, we provide a detailed comparison between different taper weight selections in the Thomson multi-taper method in the context of speaker verification. Speaker verification results on the telephone (det5) and microphone speech (det1, det2, det3 and det4) of the latest NIST 2010 SRE corpus indicate that the multitaper methods outperform the conventional periodogram technique. Instead of simply averaging (using uniform weights) the individual spectral estimates in forming the multitaper estimate, weighted averaging (using non-uniform weights) improves performance.Compared to the MFCC and PLP baseline systems, the sine-weighted cepstrum estimator 2 (SWCE) based multitaper method provides average relative reductions of 12.3% and 7.5% in equal error rate, respectively. For the multi-peak multi-taper method, the corresponding reductions are 12.6% and 11.6%, respectively. Finally, the Thomson multitaper method provides error reductions of 9.5% and 5.0% in EER for MFCC and PLP features, respectively. We conclude that both the MFCC and PLP features computed via multitapers provide systematic improvements in recognition accuracy.

This article presents a novel approach for learning domaininvariant speaker embeddings using Generative Adversarial Networks. The main idea is to confuse a domain discriminator so that is can't tell if embeddings are from the source or target domains. We train several GAN variants using our proposed framework and apply them to the speaker verification task. On the challenging NIST-SRE 2016 dataset, we are able to match the performance of a strong baseline x-vector system. In contrast to the the baseline systems which are dependent on dimensionality reduction (LDA) and an external classifier (PLDA), our proposed speaker embeddings can be scored using simple cosine distance. This is achieved by optimizing our models end-to-end, using an angular margin loss function. Furthermore, we are able to significantly boost verification performance by averaging our different GAN models at the score level, achieving a relative improvement of 7.2% over the baseline.

The 2016 edition of the NIST speaker recognition evaluation tests the ability of speaker verification systems to deal with domain mismatch between development and test data. In order to adapt to new languages, a small amount of unlabeled, in-domain data was provided-warranting the need for an unsupervised approach to learn from this data. In this work we adapt a simple domain adaptation strategy to the speaker verification problem. We test our approach using two types of speaker embeddingsi-vectors and neural network based x-vectors. Despite the simplicity of our method, we show that it outperforms a competitive PLDA domain-adaptation approach in the i-vector domain (12.11% vs 12.68% EER), and works as well in the x-vector domain (8.93% vs 8.91% EER) when the dimension of x-vectors is reduced to 150 by using LDA. Finally, as our approach adapts the speaker embeddings, we combined our adapted embeddings with the PLDA adaptation approach. We achieved our best result (8.23% EER) using this strategy with x-vector embeddings.

In this article we propose a novel approach for adapting speaker embeddings to new domains based on adversarial training of neural networks. We apply our embeddings to the task of text-independent speaker verification, a challenging, real-world problem in biometric security. We further the development of end-to-end speaker embedding models by combing a novel 1-dimensional, self-attentive residual network, an angular margin loss function and adversarial training strategy. Our model is able to learn extremely compact, 64-dimensional speaker embeddings that deliver competitive performance on a number of popular datasets using simple cosine distance scoring. One the NIST-SRE 2016 task we are able to beat a strong i-vector baseline, while on the Speakers in the Wild task our model was able to outperform both i-vector and x-vector baselines, showing an absolute improvement of 2.19% over the latter. Additionally, we show that the integration of adversarial training consistently leads to a significant improvement over an unadapted model.

No abstract

No abstract