Speaker Recognition Under Short Utterance Based On Support Vector Machine

Speaker recognition technology is a biometric method that uses speech signal torecognize different speakers. Due to its special merits in terms of flexibility，economy， accuracy and safety， speaker recognition technology owns abroadapplication future in identity verification. At present， speaker recognition systemperforms well under quiet laboratory environment and with sufficient voice. However,the performance of system will rapidly decrease when there are fewer of the speaker’svoice data in practical applications. Meanwhile, the actual application environmentcan not be predicted, resulting in the test speech and training speech do not match, andmaking the speaker recognition become more difficulty. To solve the above problem,this paper considered the system performance under the short utterance and noiserobustness.The work of this paper focuses on the method of speaker recognition under shortutterance，mainly researching on the support vector machine algorithm. Combinedwith GMM, the one-class SVM-GMM model was proposed to increase the systemperformance under short utterance. Multiple kernel mapping was introduced, themultiple kernel SVM-GMM was proposed for speaker recognition with shortutterance. For noise robustness, an extended spectral subtraction based on the cycleWiener filter and a noisy speech endpoint detection methods combination of SAP anddynamic threshold method were proposed for voice enhancement and speech endpointdetection, which can improve the quality of the input speech signal, further improvethe system performance.The main works of this paper includes:An algorithm based on one-class SVM was proposed for speaker recognitionwith short utterance. Then a new method called one-class SVM-GMM combinedone-class SVM with GMM was proposed. This method effectively reduces the timesof training and testing, consequently improve the practicality of the system.Multiple kernel mapping was introduced, the multiple kernel SVM-GMM was proposed for speaker recognition with short utterance.It combined KL function andcommon kernel function to achieve multiple kernel mapping. Used this mapping canmapped speaker’s feature to a characteristic space which combined by multiple featurespace. In the new feature space, the feature of speaker was expressed better, thus theperformance of speaker recognition system was improved.Considering the influence of noise existed in the actual application on theperformance of speaker recognition system, we used an extended spectral subtractionbased on the cycle Wiener filter for voice enhancement in the speech signalpretreatment process. This algorithm can cut noise and improve the signal-to-noiseratio of the speech signal. A noisy speech endpoint detection methods combination ofSAP and dynamic threshold method was used for endpoint detection under noiseenvironment. We combined the method mentioned above with multiple kernelSVM-GMM to further improve the performance of speaker recognition under shortutterance in noise environment.