Sparse representation classification over discriminatively learned dictionary for language recognition

Abstract

In conventional dictionary learning, the class label of the atoms is not retained. As a result of that, the location of non-zeros elements in the sparse vector (s-vector) does not infer about the true class of the test vector unlike the sparse representation classification (SRC) over the exemplar dictionary. Thus, in our earlier works employing the learned dictionary for language recognition (LR), the s-vectors are required to be processed for classification. In this work, we explore SRC over discriminative learned dictionary for LR. For this purpose, the language-specific training i-vectors are used to create the language-specific learned dictionary. These language-specific dictionaries are then concatenated to form a composite discriminative dictionary. For sparse coding, the orthogonal matching pursuit (OMP), least absolute shrinkage and selection operator (LASSO), and elastic net (ENet) algorithms are explored. The regularized multiclass logistic regression is employed for score calibration. The effectiveness of the proposed approach is validated on NIST 2009 language recognition evaluation data set in closed set condition on 30 seconds duration segments.