Language-dependent Fusion for Language Identification

A novel fusion approach for Language Identification called Languagedependent Fusion (LDF) is presented in this paper. A fusion system is a hybrid system which fuses the results from several individual sub-systems which utilize varied features, models, and/or classifiers. In LDF, instead of applying single fixed weighting coefficients to each sub-system, which happens in conventional approach such as Linear Score Weighting (LSW), varied weighting coefficients are applied to not only each sub-system but also to each language. Furthermore, instead of the experimental and statistical approach, weighting coefficients are calculated from the performance of each language-pair, which reflects the difference among languages. Experiments conducted on the OGI-92 multi-language database demonstrate a remarkable improvement when compared to individual sub-systems (45.46% error rate reduction) and commonly used fusion techniques such as LSW (33.33% error rate reduction) in a 10-language setting. Other advantages of LDF are also discussed.