| DC 欄位 |
值 |
語言 |
| DC.contributor | 電機工程學系 | zh_TW |
| DC.creator | 林立源 | zh_TW |
| DC.creator | Li-yuan Lin | en_US |
| dc.date.accessioned | 2010-6-21T07:39:07Z | |
| dc.date.available | 2010-6-21T07:39:07Z | |
| dc.date.issued | 2010 | |
| dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=975201094 | |
| dc.contributor.department | 電機工程學系 | zh_TW |
| DC.description | 國立中央大學 | zh_TW |
| DC.description | National Central University | en_US |
| dc.description.abstract | 近幾年來SVM已經被廣泛的使用在很多領域，而且有很好的效果。本論文採用NIST 2001語料庫，利用SVM來作與文字不相關的語者確認。SVM的語者確認系統，通常都是用動態核函數去處理語音自然的特性。而核函數可分為兩種，分別是參數式和微分式。本論文結合此兩種核函數來計算分數使系統的性能更好。
從UBM-MAP調適出的高斯混合語者模型參數透過GMM-supervector Kernel及Likelihood Kernel做映射，得到兩組超級向量，接著執行雜訊屬性補償(NAP)修正GMM-supervector，最後利用兩組超級向量分別訓練SVM 模型。而在仿冒者的選取上，則是選取與目標語者特徵最相似的前20名仿冒語音，使得訓練出來的SVM 模型更有鑑別力。測試時，先得到所有測試句的兩組超級向量之後，再依序分別對指定的SVM模型算分數，得到兩組分數。
從NIST 2001語料庫實驗結果顯示，64mixture的GMM-superveror Kernel (NAP)結合256mixture的derivative Kernel系統可達最好的相等錯誤率及決策成本函數分別為6.04%及0.0777，比起傳統語者確認模型的效能15.87%及0.1911，改善了9.8%及0.1135。
| zh_TW |
| dc.description.abstract | A Support vector machine-based speaker verification has become a standard approach in recent year. This thesis will evaluate the text-independent speaker verification on NIST 2001 SRE. The SVM speaker verification system usually uses dynamic kernels to handle the dynamic nature of the speech utterance. We can always separate dynamic kernel into two general classes, derivative and parametric. This paper will combine them in the score and space to promote the system’s performance.
From the UBM, we can use map to get the parameters of the GMM, by use of GMM-supervector Kernel and Likelihood Kernel to do the mapping which can get the two supervectors, and then we do the NAP process to modify the GMM-supervector. Finally we put these two supervector into the SVM for training the SVM model. About the imposters selection, we choose the top 20 speaker’s whose characteristics are similar to the target which can let the model become more discriminative. When testing, after we get all test speech’s two supervector, we use them to calculate the score with the specific model and get the score. Finally we combine two score together.
From the experiment on NIST 2001 SRE, we can find 64mixture GMM-supervector combined with a 256mixture derivative kernel result in better EER and DCF which are 6.05% and 0.1023 respectively. Comparing with the traditional speaker verification performance 15.87% and 0.1911,the proposed method obtains improvement 9.8% and 0.1135 respectively.
| en_US |
| DC.subject | 語者確認 | zh_TW |
| DC.subject | speaker verification | en_US |
| DC.title | 結合高斯混合超級向量與微分核函數之
語者確認研究 | zh_TW |
| dc.language.iso | zh-TW | zh-TW |
| DC.title | Combine GMM-Supervector and Derivative Kernel for speaker verification
| en_US |
| DC.type | 博碩士論文 | zh_TW |
| DC.type | thesis | en_US |
| DC.publisher | National Central University | en_US |