Improved DNN-HMM English Acoustic Model Specially For Phonotactic Language Recognition

Abstract

The now-acknowledged sensitive of Phonotactic Language Recognition (PLR) to the performance of the phone recognizer front-end have spawned interests to develop many methods to improve it. In this paper, improved Deep Neural Networks Hidden Markov Model (DNN-HMM) English acoustic model front-end specially for phonotactic language recognition is proposed, and series of methods like dictionary merging, phoneme splitting, phoneme clustering, state clustering and DNN-HMM acoustic modeling (DPPSD) are introduced to balance the generalization and the accusation of the speech tokenizing processing in PLR. Experiments are carried out on the database of National Institute of Standards and Technology language recognition evaluation 2009 (NIST LRE 2009). It is showed that the DPPSD English acoustic model based phonotactic language recognition system yields 2.09%, 6.60%, 19.72% for 30s, 10s, 3s in equal error rate (EER) by applying the state-of-the-art techniques, which outperforms the language recognition results on both TIMIT and CMU dictionary and other phoneme clustering methods.