Language identification through large vocabulary continuous speech recognition

In recent years, automatic language identification has become an increasingly important component in practical spoken language systems, and much attention has been devoted to various competing approaches. In this paper, we are concerned with the automatic identification of languages that may be highly similar in nature, such as the various dialects of Chinese. Our approach differs from many recent successful systems by exploiting a fusion of feature scores readily available from a large vocabulary speech recognition system. We show that such features are able to distinguish among the similar sounding dialects of Chinese, and experiments on a nine language corpus show promising performance on a three way identification task.