Duration compensation of i‐vectors for short duration speaker verification

Abstract

The standard i-vector/Gaussian probabilistic linear discriminant analysis (G-PLDA) system does not compensate for duration mismatch, which is a significant confounding factor in short duration speaker verification. A novel duration compensation technique to normalise the distribution mismatch caused by duration variation in the i-vector space is proposed. The proposed technique involves the use of two factor analysers that are tied together to share latent variables for a given speaker as the underlying generative model of the i-vector space. This leads to a transform which maps the original i-vectors onto a latent subspace that is expected to be duration invariant. The proposed method has the advantages that it normalises distribution mismatch while taking into consideration both inter-and intra-speaker variability. Experiments conducted on NIST SRE 2010 database shows that the proposed method leads to 18.54, 15.48 and 8.77% relative improvements when tested on utterances of 10, 5 and 3 s durations, respectively, compared with the best results obtained by either standard i-vector/G-PLDA or the previously proposed twin model G-PLDA.

Speaker verification with short utterances: a review of challenges, trends and opportunities

“…norm. Short Utterance Normalisation (SUN)-LDA, LDA, WCCN, CSS, Source Normalised Linear Discriminant Analysis (SN-LDA) [45] Baum Welch (BW) statistics estimation minimax [88] [91] evaluation of feature dimensionality feature dimension reduction, DKLT [40] performance comparison GMM-UBM, i-vector GPLDA [56] parallel system based on source feature M-PDSS, DCTILPR, score fusion [92] i-vector subspace projection modified-prior PLDA, score calibration, QMF [93] calibration and quality of speech signal QMFs from duration + SNR, stacked, matched/mismatched calibration 2016 [94] phonetic match between train and test WCCN, EFR, interactive voice response system [95] factor analysis on i-vector domain AFA, WCCN, LDA, GPLDA, score level fusion [96] phonetic content compensation ML-AFA, SUVN, score fusion [97] phonetic analysis modelling speech unit classes [98] normalise BW statistics compensation for feature sparsity in BW statistics [99] bootstrapped i-vectors truncate from test segment and integrating speaker similarities 2017 [100] development data with short utterance WCCN-LDA, SN-LDA, SN-WLDA, GPLDA [101] inter/intra-speaker variability a transform to map i-vectors onto a duration invariant latent subspace [102] i-vector length normalisation DNN-based length normalisation of i-vectors using PCs mentioned techniques, the information of utterance variation needs to be supplemented additionally to full-length i-vectors for modelling of PLDA. The work in [90] analysed the PLDA modelling with limited development data.…”

“…The algorithm attempted to compensate the deviation from first‐order BW statistics caused by feature sparsity and imbalance in short utterance. A compensation technique was introduced in [101] to normalise the distribution mismatch caused by duration variation in the i ‐vector space. The mentioned approach involves the use of two factor analysers, tied together, to share latent variables for a given speaker as the underlying generative model of the i ‐vector space.…”

Co-whitening of I-vectors for Short and Long Duration Speaker Verification

“…In [3], score calibration was introduced to compensate the duration mismatch. In [4] and [5], it was shown that the relationship between ivectors of short and long duration could be modeled by tying them to a single latent variable.…”