Language Identification through Parallel Phone Recognition

Language identification systems that employ acoustic likelihoods from languagedependent phoneme recognizers to perform language classification have been shown to yield high performance on clean speech. In this report, such a method was applied to language identification of telephone speech. Phoneme recognizers were developed for English, German, Japanese, Mandarin, and Spanish using hidden Markov models. Each of these processed the input speech and output a phoneme sequence in their respective languages along with a likelihood score. The language of the incoming speech was hypothesized as the language of the model having the highest likelihood. The main differences between this system and those developed in the past are that this system processed telephone speech, could identify up to five languages, and used phonetic transcriptions to train the language-specific models. The five-language, forced-choice recognition rate on 45-s utterances was 71.9%. On 10-s utterances the recognition decreased to 70.3%. In addition, it was found that adding word-specific phonemes to the training set had a negligible effect on language identification results. in j Aoosssion For i ms QRAkl öJr ifoar-rvouijoed Q Ju s t '■. '■: 1 o a t i c n ÄjT•TH s'(y'< hi^t.l m ZXOV.tlO-Qi $ ■/''vail and/or Speoial