Class-dependent two-dimensional linear discriminant analysis using two-pass recognition strategy

In this paper, we introduce a novel class-dependent extension of two-dimensional linear discriminant analysis (2DLDA) named CD-2DLDA, applied in automatic speech recognition using two-pass recognition strategy. In the first pass, the class labels of test sample are obtained using baseline recognition. The labels are then used in CD transformation of test features. In the second pass, recognition of previously transformed test samples is performed using CD-2DLDA acoustic model. The novelty of the paper lies in improvement of the present 2DLDA algorithm by its modification to more precise, class-dependent estimations repeated separately for each class. The proposed approach is evaluated in several scenarios using the TIMIT corpus in phoneme-based continuous speech recognition task. CD-2DLDA features are compared to state-of-the-art MFCCs, conventional LDA and 2DLDA features. The experimental results show that our method performs better than MFCCs and LDA. Furthermore, the results confirm that CD-2DLDA markedly outperforms the 2DLDA method.