Pair-wise language discrimination using phonotactic information

摘要

This paper describes a novel method for automatic language identification using phonotactics. Conventional phonotactic approach using N-gram language modeling requires several hours of speech data along with the corresponding orthographic transcriptions, which is not available for many of the Indian languages. This paper proposes a method which captures the language discriminating cue in co-occurance of phones using limited data. Here speech utterance is decoded into a sequence of chosen phones using an automatic phone recognizer. A unique code is assigned for each phone to obtain feature vectors corresponding to five consecutive phones. These feature vectors are then used to train a neural network / SVM based classifier at the back-end. A pair-wise language discrimination system for Hindi and Malayalam is developed using manual and automatic transcriptions.