Enhancing sparsity in linear prediction of speech by iteratively reweighted 1-norm minimization