15th Annual Conference of the International Speech Communication Association

September 14-18, 2014

Virtual Example for Phonotactic Language Recognition

Rong Tong, Bin Ma, Haizhou Li

A*STAR, Singapore

One challenge in spoken language recognition is the availability of training data. In this paper, we propose a virtual example construction method to derive artificial training examples from the existing training data. Using the proposed method, both target virtual examples and non-target virtual examples can be derived from the available training samples. An iterative virtual example selection method is proposed to select those virtual examples that may provide extra discriminative information for language separation. By incorporating virtual examples in language classifier training, the language recognition performances are improved for both closed-set and open-set tasks. Specifically, for LRE 2009 evaluation data of three durations: 30-seconds, 10-seconds and 3-seconds, the language recognition performance improved by 3.67%, 11.98%, 6.42% respectively in closed-set conditions, and 10.14%, 10.55%, 5.75% respectively in open-set conditions.

Full Paper

Bibliographic reference.  Tong, Rong / Ma, Bin / Li, Haizhou (2014): "Virtual example for phonotactic language recognition", In INTERSPEECH-2014, 3017-3021.