Identification, Based On The Language Of The Gmm-ubm Model

Language identification (LID) is the process of determining the language of a spoken utterance. It has increasing importance in speech processing applications and can be used in multi-linguistic information services and security applications.Parallel Phone Recognition-Language Model (PPR-LM) is the most classical language recognition system. However, we need lots of corpus to train phone recognizer which require manual tagging corpus. Based on GMM for language identification system does not require manual tagging corpus, so it has a good portable. especially for language recognition for minority languages in China.This thesis focus on the test-independent language identification method, using using Gaussian Mixture Model-Universal background, Language Model, Unsupervised Score Normalization to explore ways to enhance the recognition rate. The main works are as follows:(1) Building a minority-oriented language identification of telephone speech database, which includes 8 minority languages and Mandarin, each sound will have a talking time of recording and 20 questions. Addition to Chinese pronunciation of people, the words in each nation who was to use their native language and Mandarin the completion of question and answer conversation recording and automatic recording.(2) Using the principle of PPR-LM to build a new GMM-UBM-LM language recognition system. The system makes full use of acoustic information and phoneme and phonotactics. In thesis, we designed and implemented three experiments using 5 languages for the experimental data:(a) The basic GMM-UBM language identification experiments; (b) Using of acoustic scores and language model scores a direct sum of the GMM-UBM-LM Language identification experiments; (c) Using the back-end classification LDA for GMM-UBM-LM language identification experiments.(3) Utlizing Unsupervised Score Normalization to enhance GMM-UBM baseline system. This method has better performance than other methods.Experimental results show that our proposed GMM-UBM, Language Model and Unsupervised Score Normalization have better expansibility and applicability; Back-end classification (LDA) for GMM-UBM-LM system enhance 10%. Language model (LM) need more time to train and poor real time. The method of Unsupervised Score Normalization enhance 11% and better real-time systems, If it have enough testing corpora,, the system has better performance.