Citation Context

...on, we will briefly discuss computational details of this technique. Due to Weierstrass-Stone approximation theorem, polynomial classifiers are universal approximators to the optimal Bayes classifier =-=[3]-=-. Feature vectors are processed by the polynomial discriminant function. Every speaker j has w as his model, and the output of a discriminant function is j averaged over time resulting in a score for ...

Citation Context

...orage requirements. In thesliterature, there are several techniques proposed for speakersrecognition such as GMM [9], hidden Markov model (HMM)s[12], vector quantization [6] and polynomial classifier =-=[2]-=-sbased approaches, etc. In speaker identification, when a testsspeaker is predefined in speaker pool, it is referred as thesclosed-set identification corresponding to a multiple-choicesclassification ...

Citation Context

...n It has to be noted that to compare the PSOM and VQ classification techniques, only the identification without verification experiment was done. The reason for this is that the basic algorithm of VQ =-=[16]-=- does not provide a method for verification. The comparison was done on the basis of accuracy, not speed. The reason for this is that the Multiple PSOM method is aimed at high accuracy for application...

Citation Context

...frequency Cepstral Coefficients (MFCC) [8] are used as 12 dimensional input feature vectors to polynomial classifier of 2 nd or 3 rd order approximation (for speaker modeling) for all the experiments =-=[4]-=-. Feature analysis was performed using a 23.2 ms duration frame with an overlap of 50%. Hamming windows were applied to each frame and subsequently; each frame was pre-emphasized with the filter (1-0....

Citation Context

...rding and play-back of the pass phrase. Thus the other alternatives (text-independent and text-prompted), guarantee a higher security level, if used as single modality. In text-prompted systems (e.g. =-=[15]-=-) the user is required to speak a randomly selected phrase, which is not used in previous tests or during the enrolment phase. Thus it would require sophisticated hard- and software for recording and ...

Citation Context

...nty in our training data. Typically, probabilistic approaches are excellent in dealing with uncertainty, so in next stage we will try to attack the problem in a probabilistic way. [17][18][19][20][25]=-=[26]-=- 4.7 Step06: segment First, let’s try to solve the segmentation problem. At this point, to simplify the problems, we will consider the segmentation and the alignment separately. So when we consider th...

Citation Context

...uivalence) at each subband whereas T-MFCC employs Teager energy in time domain [12]. 4. Polynomial Classifier Campbell et al. first proposed polynomial classifier for speaker verification application =-=[3]-=- and later Mitra et al. applied this to speaker identification in Indian languages viz. Marathi and Hindi [11]. The basic structure of the classifier is shown in figure 3. The feature vectors are proc...

Citation Context

...hod, in term of both classification accuracy and computational complexity. 1. Introduction In the past decades, scalable algorithm have been widely used on the applications such as speech recognition =-=[1]-=-, video processing [2]. To our knowledge, there are two streamline in dealing with the large data problem. One is using hardware accelerators, such as FPGA [3], supercomputer [4]. The other is based o...