Citation Context

...each target language, to systems based on the spectral characteristics of each language usually referred to as acoustic systems. Gaussian mixture models (GMM) [1, 2] and support vector machines (SVM) =-=[3]-=- have been the classifiers of choice over recent years for acoustic modeling, consistently outperforming their high-level counterparts. For the GMM and SVM classifiers, a number of techniques develope...

Citation Context

...ghtly outperform WLP in speaker recognition. In speaker recognition, it is common to fuse FFT- and LPderived features since that they capture complementary properties of the underlying speech process =-=[22, 23]-=-. Here, we consider fusion of the FFT- and SWLP-based features using two well-known fusion strategies. Score fusion is carried out by summing the log-likelihood ratio scores of the two classifiers, sc...

Citation Context

...ori knowledge (world) and the current stat, copy it into client model. 2.2.5. Discriminant classifiers Discriminant classifiers like the SVM were proposed during the past years in several works as in =-=[10, 11, 12, 13]-=-. These classifiers are usually applied to GMM supervectors. A GMM supervector is composed of the means of a classical GMM system, as initially proposed by [13]. Libsvm library 3 is used for the basic...

Citation Context

...erability of GMM recognizer against voice conversion attacks also on the SRE 2006 corpus. In the present study we carry out more thorough comparison including five speaker recognizers. Three of these =-=[15, 16, 17]-=- – used for reference purposes – are lightweight recognizers without intersession compensation or external score normalizations. The other two, GMM supervector [18] with nuisance attribute projection ...

Citation Context

...he-art performance as judged by the latest NIST SREs. The realization has taken place in the context of the open-source ALIZE toolkit [2], [3]. Key developments include support vector machines (SVMs) =-=[4]-=-, [5], the associated nuisance attribute projection compensation (NAP) [6], and factor analysis (FA) [7], [8]. We trace individually their influences on ASV performance. As a starting point we choose ...

Citation Context

... worth noting that when the base models used to define a generative kernel are identical (i.e. the parameters are tied, λ (1) =λ (2) ), a standard Fisher kernel is obtained. 3.2.3 Sequence kernels In =-=[12]-=- sequence kernels were proposed. Unlike the kernels discussed thus far, sequence kernels separate the definition of the high-dimensional feature-space from the handling ofsCHAPTER 3. DYNAMIC KERNELS 3...

Citation Context

... utterance O v can be scored using: S(O v ,s) = N∑ i=1 α (s) i yiK(Oi,O v ) + b (s) The nature of the dynamic kernel K(Oi,Oj) depends upon the form of φ(·). Dynamic kernels based upon vectoraveraging =-=[14]-=- or Dynamic Timewarping [15] have been developed but these have now been largely superseded by gener(9)LONGWORTH AND GALES: COMBINING DERIVATIVE AND PARAMETRIC KERNELS FOR SPEAKER VERIFICATION 3 ativ...

Citation Context

... (1) can be split as follows, f(x) = ∑ αiK(x, xi)− ∑ αiK(x, xi)+d. (13) {i|αi>0} {i|αi<0} The two terms on the right hand side resemble a log-likehood ratio between an in-class and out-of-class model =-=[12]-=-. Since we are working with a linear kernel, we can write (13) as f(x) = w t pDx − wt mDx + d. (14) where D is a diagonal matrix and wp = ∑ αixi, wm = ∑ {i|αi>0} {i|αi<0} αixi. (15) For the kernel (12...

Citation Context

...MMs on several text-independent speaker verification tasks. One of these systems, based on an explicit polynomial expansion (Campbell, 2002) has obtained good results during the NIST 2003 evaluation (=-=Campbell et al., 2006-=-), but suffers from a lack of theoretical interpretation and justification. Moreover the approach precludes the use of the so-called kernel trick, which is at the heart of the flexibility of SVM based...