Mahalanobis Metric Scoring Learned from Weighted Pairwise Constraints in I-Vector Speaker Recognition System

Abstract

The i-vector model is widely used by the state-of-the-art speaker recognition system. We proposed a new Mahalanobis metric scoring learned from weighted pairwise constraints (WPCML), which use the different weights for the empirical error of the similar and dissimilar pairs. In the new i-vector space described by the metric, the distance between the same speaker's i-vectors is small, while that of the different speakers' is large. In forming the training set, we use the traditional way in random fashion and develop a new nearest distance based way. The results on the NIST 2008 telephone data shown that our model can get better performance than the classical cosine similarity scoring. When using the nearest distance based way to form the training set, our model is better than the state-of-the-art PLDA. And the results on the NIST 2014 i-vector challenge show that our model is also better than the PLDA.

Emotional Variability Analysis Based I-Vector for Speaker Verification in Under-Stress Conditions

“…Another study reported [32] that Cosine or Euclidean scoring methods provide a significant improvement than PLDA. The effectiveness of the Mahalanobis scoring method has been explored by [37,38] and presented an excellent performance for the i-vector system in the speaker recognition system. In this paper, we assess the effectiveness of the speaker verification system in different scoring methods, such as Cosine similarity scoring (CSS), Euclidean distance scoring (EDS), and Mahalanobis distance scoring (MDS).…”

“…The effectiveness of the Mahalanobis metric for speaker detection scoring has been proven by [37,38]. The score between two i-vectors w target and w test is proportional to the log-probability that both i-vectors belong to a unique class following the covariance matrix τ.…”

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

“…The development data with short speech segments can be used to train the metric, and subsequently the learned metric can be used for measuring the similarity of i ‐vectors in a pairwise manner. Previous studies in distance metric method in SV do not focus on short‐utterance problem [116, 117]; however, the approach could be explored further to investigate the similarity between i ‐vectors of small speech segments. Sparse methods: The limited data conditions in SV lead to sparsity in sufficient statistics estimation which is successively used in i ‐vector estimation [48, 98]. Methods developed to handle the sparsity issue such as dictionary learning [118], sparse representation [119] can be investigated to effectively process and represent the speech data for short utterances.…”

“…The development data with short speech segments can be used to train the metric, and subsequently the learned metric can be used for measuring the similarity of i ‐vectors in a pairwise manner. Previous studies in distance metric method in SV do not focus on short‐utterance problem [116, 117]; however, the approach could be explored further to investigate the similarity between i ‐vectors of small speech segments.…”