UTD-CRSS Systems for 2016 NIST Speaker Recognition Evaluation

Zhang, Chunlei; Bahmaninezhad, Fahimeh; Ranjan, Shivesh; Yu, Chengzhu; Shokouhi, Navid; Hansen, John H. L.

doi:10.21437/interspeech.2017-555

Cited by 11 publications

(8 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, i-vector systems are prone to have performance degradation when short utterances are met in enrollment/test phase. Fig.1 is the DET curves with respect to different durations of test utterances in CRSS submissions for SRE16 [10]. A clear speaker verification performance drop can be found in this analysis.…”

Section: Introductionmentioning

confidence: 70%

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

Zhang

Koishida²

2017

Interspeech 2017

Self Cite

214

135

View full text Add to dashboard Cite

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.

show abstract

Section: Introductionmentioning

confidence: 70%

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

Zhang

Koishida²

2017

Interspeech 2017

Self Cite

214

135

View full text Add to dashboard Cite

show abstract

“…Thereafter, i-vectors are post-processed with lengthnormalization and LDA. Eventually, PLDA is trained and final log-likelihood scores are calculated [28].…”

Section: Speaker Recognition Evaluationmentioning

confidence: 99%

Convolutional Neural Network Based Speaker De-Identification

Bahmaninezhad¹,

Zhang²,

Hansen³

2018

The Speaker and Language Recognition Workshop (Odyssey 2018)

Self Cite

View full text Add to dashboard Cite

Concealing speaker identity in speech signals refers to the task of speaker de-identification, which helps protect the privacy of a speaker. Although, both linguistic and paralinguistic features reveal personal information of a speaker and they both need to be addressed, in this study we only focus on speaker voice characteristics. In other words, our goal is to move away from the source speaker identity while preserving naturalness and quality. The proposed speaker de-identification system maps voice of a given speaker to an average (or gender-dependent average) voice; the mapping is modeled by a new convolutional neural network (CNN) encoder-decoder architecture. Here, the transformation of both spectral and excitation features are studied. The voice conversion challenge 2016 (VCC-2016) database is used to train the system and examine performance of the proposed method. We use two different approaches for evaluations:(1) objective evaluation: equal error rates (EERs) calculated by an i-vector/PLDA speaker recognition system range between 1.265 -3.46 % on average for all developed systems, and (2) subjective evaluation: achieved 2.8 naturalness mean opinion score (MOS). Both objective and subjective experiments confirm the effectiveness of our proposed de-identification method.

show abstract

“…However, these systems often rely on a large collection of in-domain and well-annotated data, e.g., transcriptions for ASR DNN acoustic modeling and speaker labels for PLDA training [5,6]. Studies have shown a significant performance gap between in-domain and out-of-domain systems [7,8,9,10]. Also, it is expensive to collect a large amount of labeled data for every new domain.…”

Section: Introductionmentioning

confidence: 99%

“…However, it is noted that the DAC corpus is mostly English speech. Consequently, we find that similar techniques are not so effective in the NIST SRE16 setup, where a more severe domain mismatch (i.e., language mismatch between training data and enrollment/test data) is designed to encourage effective domain adaptation methods [9,10]. This This project was funded in part by AFRL under contract FA8750-15-1-0205 and partially by the University of Texas at Dallas from the Distinguished University Chair in Telecommunications Engineering held by J. H. L. Hansen.…”

Section: Introductionmentioning

confidence: 99%

An Analysis of Transfer Learning for Domain Mismatched Text-independent Speaker Verification

Zhang¹,

Ranjan²,

Hansen³

2018

The Speaker and Language Recognition Workshop (Odyssey 2018)

Self Cite

View full text Add to dashboard Cite

In this paper, we present transfer learning for deep neural network based text-independent speaker verification, in the presence of a severe mismatch between the enrollment and the test data. Given a pre-trained speaker embedding network developed with out-of-domain data, we explore and analyze how this pre-trained model can benefit for the in-domain speaker verification task. Two alternative strategies are investigated to perform transfer learning, i.e., vanilla transfer learning (V-TL) and curriculum learning based transfer learning (CL-TL). The proposed methods are validated on UT-SCOPE-physical speech corpus, where we create a setup to introduce mismatched evaluation conditions with the neutral and the physical task stressed speech. Experimental results confirm the effectiveness of both V-TL and CL-TL techniques. Employing transfer learning based on the pre-trained model, we are able to achieve a +47.7% relative improvement over a conventional i-vector/PLDA system and a +30.6% relative improvement over a recent proposed end-to-end system, respectively.

show abstract

UTD-CRSS Systems for 2016 NIST Speaker Recognition Evaluation

Cited by 11 publications

References 9 publications

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

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

Convolutional Neural Network Based Speaker De-Identification

An Analysis of Transfer Learning for Domain Mismatched Text-independent Speaker Verification

Contact Info

Product

Resources

About