Improving GMM–UBM speaker verification using discriminative feedback adaptation

Abstract

The Gaussian Mixture Model -Universal Background Model (GMM-UBM) system is one of the predominant approaches for text-independent speaker verification, because both the target speaker model and the impostor model (UBM) have generalization ability to handle "unseen" acoustic patterns. However, since GMM-UBM uses a common anti-model, namely UBM, for all target speakers, it tends to be weak in rejecting impostors' voices that are similar to the target speaker's voice. To overcome this limitation, we propose a discriminative feedback adaptation (DFA) framework that reinforces the discriminability between the target speaker model and the anti-model, while preserving the generalization ability of the GMM-UBM approach. This is achieved by adapting the UBM to a target speaker dependent anti-model based on a minimum verification squared-error criterion, rather than estimating the model from scratch by applying the conventional discriminative training schemes. The results of experiments conducted on the NIST2001-SRE database show that DFA substantially improves the performance of the conventional GMM-UBM approach.1

An Algorithm of Speaker Clustering Based on Model Distance

“…2). GMM-UBM can describe the differences between speakers from the point of statistics and it can generate more sophisticated model description under a less speech training set, which has better robustness [15][16].…”

“…It employs every speech segment for clustering according to MAP (Maximum A Posteriori), obtains GMM (Gaussian Mixture Model) of every speech segment for clustering adaptively, then constructs affinity matrix of spectral clustering by calculating probability distance based on a finite observation sequence model. GMM-UBM (Gaussian Mixture Model -Universal Background Model) is used in this paper can describe the speakers' differences from the point of statistics adequately, GMM-UBM can produce more precise model expression under the condition of a small voice training set, which has better robustness [15][16]. Probability distance based on a finite observation sequence model in this paper considers not only the models differences but also samples, and it can overcome the shortcoming of model description that is not accurate enough under the condition of the relative short length of speech segment.…”

Discriminative Universal Background Model Training for Speaker Recognition

“…Unlike sigmoid function, the squared loss function has greater gradient for large d, which gives more penalty for the severe false verification segments. To give an intuitive illustration, we borrow a figure from (Chao et al, 2009) and show it in Fig. 3.…”

“…3. (Chao et al, 2009) Using this squared loss function, the gradient of the objective function w.r.t the mean vector will be…”

“…In (Angkititrakul & Hansen, 2007), the training process is divided into two stages: in the first stage, the MCE is used minimize the classification error among the in-set speaker models; in the second stage, the MVE is used to minimize the verification error between the in-set and background models. In (Chao et al, 2008;2009), the MVE methods are used to reinforce the discriminability between the target speaker model and the target speaker dependent anti-model. (2) Other approaches attempt to discriminatively adapt the target speaker model from the UBM, which can be viewed as the modification of the classical maximum a posteriori (MAP) adaptation (Gauvain & Lee, 1994).…”