Language identification using PPRLM with confidence measures

In this paper, confidence measures are incorporated into the traditional language identification system - parallel phone recognizers followed by language modeling (PPRLM). This new system can be applied in open-set language identification environments. Two confidence measure tactics are tested. One with a composite background language model explicitly and the other is using online garbage model. Three different computing methods of confidence measures are used to score the recognition results. Experimental results show that the system with composite background language model outperforms the system without background model, and using confidence measures in the final decision can greatly improve the performance.