On the Use of Phonotactic Vector Representations with FastText for Language Identification

This paper explores a better way to learn word vector representations for language identification (LID). We have focused on a phonotactic approach using phoneme sequences in order to make phonotactic units (phone-grams) to incorporate context information. In order to take into consideration the morphology of phone-grams, we have considered the use of sub-word information (lower-order n-grams) to learn phone-grams embeddings using FastText. These embeddings are used as input to an i-Vector framework to train a multiclass logistic classifier. Our approach has been compared with a LID system that uses phone-gram embeddings learned through Skipgram that do not implement sub-word information, using Cavg as a metric for our experiments. Our approach to LID to incorporate sub-word information in phone-grams embeddings significantly improves the results obtained by using embeddings that are learned ignoring the structure of phone-grams. Furthermore, we have shown that our system provides complementary information to an acoustic system, improving it through the fusion of both systems.