Language-Dependent Contribution Measuring and Weighting for Combining Likelihood Scores in Language Identification Systems

Developing a fusion-based system is one of the key research issues in modern Language Identification (LID) systems. In this paper we investigate existing fusion techniques for LID systems and propose an alternative solution. By directly utilizing language-dependent contribution information, a novel Language-Dependent Weighting approach is introduced and implemented. We investigate various contribution measures, including LID performances, likelihood ratios, and Kullback–Leibler divergence. These measures are conducted from either development datasets or class models. The advantage of using language-dependent weighting over language-independent weighting is illustrated using a Language-Dependent Contribution Map. Both the OGI and CallFriend databases show a very similar contribution pattern which is related to language characteristics. Experiments on the NIST LRE 2003 task and OGI database demonstrate that the proposed fusion technique outperforms other recent fusion techniques when the amount of available development data is limited. In particular, the system based on Kullback-Leibler divergence achieved the best performance while eliminating the need for development data.