A Study on Low-resource Language Identification

Abstract

Modern language identification (LID) systems require a large amount of data to train language-discriminative models, either statistical (e.g., i-vector) or neural (e.g., x-vector). Unfortunately, most of languages in the world have very limited accumulation of data resources, which result in limited performance on most languages. In this study, two approaches are investigated to deal with the LID task on low-resource languages. The first approach is data augmentation, which enlarges the data set by incorporating various distortions into the original data; and the second approach is multi-lingual bottleneck feature extraction, which extracts multiple sets of bottleneck features (BNF) based on speech recognition systems of multiple languages. Experiments conducted on both the i-vector and x-vector models demonstrated that the two approach are effective, and can obtain promising results on both in-domain data and out-of-domain data.