Research And Implementation On Language Recognition System Based On GSV-SVM

Language recognition is the process of distinguishing which language a given utterancebelongs to by means of analysis and processing of computer, which is an important research directof speech recognition. With the acceleration of globalization, language recognition has a widelyapplication prospect in multilingual information services, machine translation and militarysecurity.This dissertation relies on a key project of the National High Technology DevelopmentProgram of China (863Program) in the field of information technology, whose purpose is todevelop the language recognition system which can process multi-access voices real-timely.Based on the in-depth analysis of the state of the art, this dissertation studies the GaussianMixture Model SuperVector-Support Vector Machine (GSV-SVM) based language recognitionsystem, with emphasis on mismatch between train and test conditons, improving the performanceof short-duration in language recognition and implementation of system on multicore DSP. Themain work and innovations obtained in this dissertation are outlined as follows:1. The key techniques of the language recognition system based on GSV-SVM are studied.Aiming at the matter of the feature parameters of language recognition being liable to be affectedby the channel and speaker, compensation techniques in feature domain are studied and discussedby experiments, whose results are used to arrange the reasonable order of techniques as thedefault configuration of the baseline system. The baseline system achieves good performance,laying the foundation down for the latter studies.2. Discriminating Weighted Nuisance Attribute Projection is proposed. To deal with thematter of mismatch between train and test conditions, the proposed algorithm quantitativelyestimates the source of nuisance based on scatter of the test utterances’s eigenvalues of givenlanguage. The normalized estimates are used as discriminating weight in the training of projectionmatrix. The projection matrix of this algorithm can more thoroughly remove the nuisance of thechannel and the speaker independent of language. This algorithm simplifies the training processof projection matrix and reduces the requirement of label information of training data.Experimental results show that the proposed algorithm can effectively improve the recognitionperformance of the system.3. Discriminative Model Pushing is proposed. To improve the performance ofshort-duration in language recognition, the proposed algorithm moves the support vectors in thedirection of the normal to the separation hyperplanes trained by SVM. Then the moved supportvectors are pushed back to reconstruct GMMs of the target language and non-target language,which are used to calculate the logarithmic likelihood score of feature parameters for outputjudgment. This algorithm retains the advantage of GMM in short utterances and the welldiscriminative information of GSV-SVM. The model can adequately describe the distribution offeature parameters. Feature domain Discriminating Weighted Nuisance Attribute Projection isproposed to remove the nuisance of feature parameters independent of language, based on the characteristic of Discriminative Model Pushing in judgment. Experimental results show that theproposed algorithm can evidently improve the recognition performance of the system, with moreeffectively improvement on10sec’s evaluation over30sec’s evaluation.4. Language recognition system based on TMS320C6678is implemented. Based on thecharacteristics of the TMS320C6678platform, the improved algorithms and codes of languagerecognition are optimized. Considering the architecture of language recognition system, tasks aredesigned in parallel with effective configuration of the computing resources and storage resourceson the platform. Experimental results show that the language recognition system based onTMS320C6678is able to fulfill the132-access recognition tasks at least and the performance isalmost the same with the result from VC++2010, which provide reliable guarantee formulti-access real-time processing and accurate recognition of the language recognition system.