Unifying Probabilistic Linear Discriminant Analysis Variants in Biometric Authentication

Abstract

Abstract. Probabilistic linear discriminant analysis (PLDA) is commonly used in biometric authentication. We review three PLDA variants -standard, simplified and two-covariance -and show how they are related. These clarifications are important because the variants were introduced in literature without argumenting their benefits. We analyse their predictive power, covariance structure and provide scalable algorithms for straightforward implementation of all the three variants. Experiments involve state-of-the-art speaker verification with i-vector features.

Intra-class covariance adaptation in PLDA back-ends for speaker verification

“…While there are several variants of the PLDA algorithm [17], a generic implementation models the interclass and intraclass variances as follows…”

Variants of mel-frequency cepstral coefficients for improved whispered speech speaker verification in mismatched conditions

“…Matching between a test utterance and a target speaker can be done using either a fast scoring technique based on cosine distance or on probabilistic linear discriminant analysis (PLDA) [12], [14] based scoring. The PLDA model splits the total data variability into within-individual and betweenindividual variabilities, both residing in small-dimensional subspaces.…”

“…The PLDA model splits the total data variability into within-individual and betweenindividual variabilities, both residing in small-dimensional subspaces. Originally introduced for face recognition, PLDA has become a standard in speaker recognition, and details can be found in [14]. Using φ to denote the i-vector extracted from a given speech recording, in a verification scenario, there are two possible hypotheses: 1) φ test and φ enrol share the same class, and 2) φ test and φ enrol are from different classes.…”

“…Using φ to denote the i-vector extracted from a given speech recording, in a verification scenario, there are two possible hypotheses: 1) φ test and φ enrol share the same class, and 2) φ test and φ enrol are from different classes. Lastly, the corresponding score can be obtained by computing the log-likelihood between the two hypotheses, which is given by s = ln(P (φ test , φ enrol )) − ln(P (φ test )P (φ enrol )); details can be found in [14]. For the experiments herein, the open-source Bob signal processing toolbox was used [15].…”

Joint Speaker Verification and Antispoofing in the <inline-formula> <tex-math notation="LaTeX">$i$ </tex-math></inline-formula>-Vector Space

“…If we go even further and set both subspace matrices V and U to have the full rank we get so called two-covariance model [35]. Comparison of all three PLDA models as well as the EM-algorithms to train them are presented in [36].…”