Improvement of language Identification performance by Aggregated Phone Recognizer

Two popular and better performing approaches to language Identification (LID) are Phone Recognition followed by Language Modeling (PRLM) and Parallel PRLM. In this paper, a new LID approach named Aggregated PRLM or APRLM is proposed. In PRLM based LID systems, only one phone recognizer is used, independently of the language targets. At the opposite, in PPRLM based LID systems, multiple phone recognizers are used, but always independently of the language targets. So it may happen that all phones of a language target don't occur in at least one of the tokenizers provided by the phone recognizers. In this paper, it is proposed that after the phone recognition step, to aggregate the phone sequences obtained by multiple phone recognizers and to provide a new phone sequence. Several language identification experiments were conducted and the proposed improvements were evaluated using OGI-MLTS corpus. Our results show that APRLM overcomes PPRLM about 1.3% in two language classification tasks.