Unique n-Phone Ranking Based Spoken Language Identification

This paper presents a novel approach to phonetic-based language identification (LID). Motivated by the assumption underlying phonotactic LID that accounting for permissible phone sequences supports the process of distinguishing one language from another, this paper presents a novel approach based on the automatic identification of phone sequences of different lengths unique to a language, which are subsequently employed to determine which language has been spoken. The approach implements a discriminative training that involves a ranking procedure on unique phone sequences generated by a single-language phone recogniser. As a baseline system, phone recognition followed by language modelling (PRLM) is used to compare the performance of the proposed approach. Having 21 languages in the target set, the experiments show that the proposed approach achieves better accuracies. Moreover, it requires a 5.5 times shorter processing time than the baseline system.