On the use of i–vector posterior distributions in Probabilistic Linear Discriminant Analysis

Abstract

The i-vector extraction process is affected by several factors such as the noise level, the acoustic content of the observed features, the channel mismatch between the training conditions and the test data, and the duration of the analyzed speech segment. These factors influence both the i-vector estimate and its uncertainty, represented by the i-vector posterior covariance. This paper presents a new PLDA model that, unlike the standard one, exploits the intrinsic i-vector uncertainty. Since the recognition accuracy is known to decrease for short speech segments, and their length is one of the main factors affecting the i-vector covariance, we designed a set of experiments aiming at comparing the standard and the new PLDA models on short speech cuts of variable duration, randomly extracted from the conversations included in the NIST SRE 2010 extended dataset, both from interviews and telephone conversations. Our results on NIST SRE 2010 evaluation data show that in different conditions the new model outperforms the standard PLDA by more than 10% relative when tested on short segments with duration mismatches, and is able to keep the accuracy of the standard model for long enough speaker segments. This technique has also been successfully tested in the NIST SRE 2012 evaluation.

Incorporating uncertainty as a Quality Measure in I-Vector Based Language Recognition

“…Generative and discriminative models are two general approaches for language recognition based on i-vectors. Although the reported results using discriminative methods such as multiclass logistic regression and support vector machines are comparable to those of using generative models [22] such as Gaussian and probabilistic linear discriminant analysis (PLDA) models, the generative models provide an appropriate framework to benefit from the uncertainty in the i-vector extraction process through the posterior covariance matrix of the i-vector [23]. PLDA [24], originally studied in image processing, has been very successful in speaker and language recognition.…”

“…Fig.1 indicates a high correlation (0.98) between the proposed quality measure and utterance duration in the LRE15 database (described in Section 5.1). However, since the i-vector posterior covariance is also influenced by other factors such as background noise, channel type, incomplete transformations and the acoustic content of the utterance [19,23], we expect that the proposed quality measure captures more information about the quality of an utterance than its duration.…”

“…In contrast, length normalization [30] that maps i-vectors on the unit sphere bỹ φ = φ/ φ does not satisfy the Gaussian distribution and hence, the Gaussian assumption in PLDA is no longer applicable. To address this issue, one can either use a non-Gaussian assumption for the PLDA model such as the heavy-tailed PLDA model [25] or make the transformation linear using first-order Taylor series expansion around the i-vector posterior mean [23]. Applying a simplified version of first-order Taylor expansion around the i-vector posterior mean [23] results in the length normalized i-vector posterior mean and covariance as:…”

An Improved Uncertainty Propagation Method for Robust I-vector Based Speaker Recognition

“…Other work based on the aleatoric uncertainty concept has also given rise to uncertainty propagation approaches for speaker recognition. However, these approaches focused on the issue of computing representations with in-sufficient data, caused by utterances with different, possibly short durations [9][10][11][12][13][14][15].…”

End-to-End Residual CNN with L-GM Loss Speaker Verification System

“…The conventional speaker verification approach entails using i-vectors [3] and probabilistic linear discriminant analysis (PLDA) [2]. As a supervised learning method, i-vector requires sufficient statistics which are computed from a Gaussian Mixture Model-Universal Background Model (GMM-UBM), followed by a PLDA model to produce verification scores [3].…”