Language model adaptation and confidence measure for robust language identification

This paper describes two methods to improve the robustness of the language identification system in practical applications. One is a language model adaptation method, which modifies the language model parameters automatically to solve the mismatch problem in different channels. And the other is a confidence measure based method, which proves to be more effective comparing to conventional score based method. Experiments show that with the use of these two methods, the performance of system is greatly improved. Tested on the MCTS (multi-channel telephone speech) database, the average error rate decreases from 15.81% to 12.92% for the baseline.