A Robust Equalization Feature for Language Recognition

The performance of language recognition system is mainly determined by feature extraction and model training. In this paper, a robust equalization feature for language recognition is proposed, which utilizes the common features of the speech spectrum mean vector to calculate a global mean vector. The spectrum mean vector of each segment is equalized on the global mean vector, and the equalization features are obtained. In model training, Gated Recurrent Unit (GRU) of Recurrent Neural Network (RNN) is applied to language recognition, in which GRU can reduce the amount of computation and shorten the training time. The experimental results show that the proposed method outperforms the baseline system on the NIST LRE 2007 corpus.