Automatic spoken language identification (LID) is the task of determining the language identity of a particular speech segment. Token-based approaches have proved quite effective for LID tasks. A traditional token-based technique is phone recognition followed by language modeling (PRLM), which employs n-gram language models to derive phonotactic scores.

Recently, researchers of Thinkit Speech Lab, Institute of Acoustics, Chinese Academy of Sciences carried out a series of studies and created a novel token-based approach for LID by using bayesian logistic regression model. The new approach takes into account prior distribution for parameters of logistic regression models in order to avoid overfitting.

The researchers first decode the speech utterances into token sequences, and then they design a hierarchical system which utilizes bayesian logistic regression model to perform LID task on these token sequences. Experiments are conducted on the NIST LRE 2007 database and the practical results show that the proposed approach exhibits quite competitive performance and proves to be complementary with other state-of-the-art token-based approaches.

The research result was published on the recently issued journal of IEICE Electronics Express (Volume 7, Issue 6, March 2010, Pages 390-396).