Sparse Representation Classification based Language Recognition using Elastic Net

In our earlier work, we have explored the sparse representation classification (SRC) for language recognition (LR) task. In those works, the orthogonal matching pursuit (OMP) algorithm was used for sparse coding. In place of l(0)-norm minimization in the OMP algorithm, one could also use l(1)-norm minimization based sparse coding such as the least absolute shrinkage and selection operator (LASSO). Though leading to better sparse representation, the LASSO algorithm is quite latent in contrast to the OMP. This work explores the elastic net (ENet) sparse coding algorithm in SRC based LR framework. Unlike conventional sparse coding methods, the ENet employs both l(1) and l(2) constraints in regularizing the sparse solutions, thus is expected to yield improved sparse coding. The experiments are performed on NIST 2007 LRE data set in closed set condition on 30 seconds duration segments. Scores are calibrated using regularized multi-class logistic regression. For language representation, the utterances are mapped to the well-known i-vector representation and applied with the within-class covariance normalization (WCCN) based session/channel compensation. The proposed ENet based LR approach is noted to significantly outperform the other LR methods developed using existing sparse and non-sparse representations.