Automatic text independent language identification using reduct set of feature vectors

In this paper, robust features are proposed for spoken language identification (LID) system. 12 Mel frequency cepstral coefficients (MFCCs) and five formant frequencies are extracted from each short-time windowed speech signal. These features are concatenated to form 17-dimensional feature vectors. 8-dimensional reduct set is obtained from this 17-dimensional feature vector using rough set theory. This 8-dimensional reduct set is transformed into 15 dimensional new feature vectors using the approach followed in [1]. In both the training and testing phases of LID, these 15 dimensional feature vectors are used. Due to usage of reduct set, there is a significant reduction of time in training and testing phases of the proposed LID system. The experiments are carried out on speech database of Indian languages and the results are impressive.