Automatic turn-level language identification for code-switched Spanish–english dialog

Abstract

We examine the efficacy of text and speech-based features for language identification in code-switched human-human dialog interactions at the turn level. We extract a variety of character- and word-based text features and pass them into multiple learners, including conditional random fields, logistic regressors and deep neural networks. We observe that our best-performing text system significantly outperforms a majority vote baseline. We further leverage the popular i-Vector approach in extracting features from the speech signal and show that this outperforms a traditional spectral feature-based front-end as well as the majority vote baseline.