A study of feature selection in phonotactic language recognition
Chunyan Liang, Lin Yang, Junjie Wang, Yonghong Yan · The Journal of the Acoustical Society of America · 2012
This paper is a comparative study of feature selection methods in phonotactic language recognition. The phonotactic feature is presented by n-gram statistics derived from one or more phone recognizers in the form of high dimensional feature vectors. Feature selection is necessary for its ability of reducing the dimension of feature vectors so that the higher order n-gram features can be adopted in language recognition. This paper investigates four feature selection strategies that are introduced from text categorization, including mutual information (MI), Chi-squared test (CHI), information gain (IG) and weighted log likelihood ratio (WLLR). These methods are compared on the NIST 2009 Language Recognition Evaluation (LRE) task. The experimental results show that CHI, IG and WLLR can effectively obtain much lower dimensional features without affecting the language recognition performance. In contrast, MI has relatively poor performance due to its bias towards favoring rare terms. This work is partially supported by the National Natural Science Foundation of China (No. 10925419, 90920302, 10874203, 60875014, 61072124, 11074275, 11161140319).