The Study Of GMM-based Language Identification

Language identification has been rapid developed in the past 10 years. It comes from theory to practical application. The aim of this thesis is to construct the GMM-based language identification system framework, and to minimize the system's error rate on NIST LRE test corpus. The detailed research and result of this paper is abstracted as follows:Firstly, this paper proposed the GMM-UBM based framework as our baseline. Our baseline system mainly has three modules: feature extraction module, model training module and evaluation module. Feature extraction module extract feature which diminish the variability of channel, speaker and so on. We employs a number of techniques that have proven to improve GMM modeling capability, such as Vocal Tract Length Normalization, RelAtive SpecTrAl filtering, Shifted Delta Cepstra. Model training module iteratively trains UBM and each language acoustic model with Maximum Likelihood Estimation criterion. Evaluation module process scores with Gaussian backend classifier, which improve the performance of system remarkable. In NIST LRE corpus, our baseline system achieves a good performance.Secondly, factor analysis has been proposed. In GMM-based language identification system, the variability of the speaker and channel is one of the most important factors affecting the performance. Session variability subspace is estimated based with EM algorithm on the property of language identification. Both model and feature domain compensation methods are proposed. In NIST LRE 2007 corpus 30s evaluation, the equal error rate (EER) of the proposed system can reduce by 36.5% against the baseline GMM system。Lastly, Discriminative training of GMM models with Maximum Mutual Information criterion is discussed. At first, a lot of exam is done in MMI training, such as: the training utterance length, utterance segmentation of hard decision or phoneme recognizer decision, TOP N stratage for time consumption problem, the problem of dialect in MMI. Next, both discriminative training and factor analysis methods are used in our system. In NIST LRE 2007 corpus 30s evaluation, EER equals 2.13%. Our system achieves state-of-the-art performance in GMM area.