Data-Driven Phone Selection for Language Identification via Bidirectional Long Short-Term Memory Modeling

Abstract

In this paper, we propose a new phone selection method to select more suitable phones with higher score for language identification (LID), which is more similar to target language. A data-driven approach is developed for the phone selection to avoid using complex semantic knowledge which benefits from significant reduction in the manual cost of learning different languages. Recently, bidirectional long short-term memory (BLSTM) can provides more accurate content frame alignments with sequence information from longer duration, which has improved automatic speech recognition (ASR) performance. In principle, the output of BLSTM based ASR contains more candidates in form of phone lattice, which can reduces adverse effect of many practical factors, such as variations of channels, noises and accents. Therefore, initial phones sequences are extracted from phone lattice firstly which are generated by speech recognition results of BLSTM based ASR system. Second, asymmetrical distance between each phone and target language is proposed and then applied to weight the initial phones sequences. Accordingly, language-related phones are selected from the weighted phones. Finally, the selected phones are used to re-score input sentences for the LID system. Intensive experiments have been conducted on AP16-OLR Challenge to validate the effectiveness of our proposed method. It can be seen from results, these selected phones are more effective to LID than the rest phones. Our method gives improvement up to 39.96% in terms of C_avg compared with method without using phone selection.