Citation Context

...ng window of 3 s [14], [17]. The latter method had previously been found by one study to be more effective than other standard normalization techniques on a speaker verification task on cellular data =-=[45]-=-. In [17], it was found the feature normalization was necessary to get significant gain from the cluster recombination technique. When the clusters are merged, a new speaker model can be trained with ...

Citation Context

...hort-time Gaussianization (e.g., [10]–[13]). Score-domain compensation aims to remove handset-dependent biases from the likelihood ratio scores. The most prevalent methods include H-norm [14], Z-norm =-=[15]-=-, and T-norm [16]. Examples of the model-domain compensation methods include the speaker-independent variance transformation [17], and the transformation for synthesizing supplementary speaker models ...

Citation Context

...f 0-3.8 Khz and a framerate of 10ms. Observations were comprised of 15 static and 15 delta mel-PLP coefficients and the delta energy. Static energy coefficients were not included since previous works =-=[26]-=- have shown that they contain no speakerdiscriminant information. To introduce additional robustness to noise, Cepstral Mean Subtraction was performed followed by Cepstral Feature Warping [27] using a...

Citation Context

...(e.g., credit card information, bank account balance, etc) or other sensitive customer information (e.g., healthcare records). Examples include work by Bimbot et al [5] and Lamel, Barras, and Gauvain =-=[6, 7]-=-. One of the challenges in implementing speaker verification on handheld devices arises from their greatest attribute: mobility. Unlike vestibule security systems, handheld devices are often used in h...

Citation Context

...mprovement. This choice is also made in [24], where intersession variability is of interest. Cepstral mean and variance normalization was performed afterwards. Following the system setup described in =-=[25]-=-, we trained gender-dependent UBMs and investigated the effect of the Tnorm score normalization [22] on system performance. For each target speaker, a speaker-specific GMM with diagonal covariance mat...

Citation Context

...n results by about 10%. In this paper, we demonstrate that verification results can be boosted by 50% when cohort normalization is applied in conjunction with image-set matching. Specifically, Z-norm =-=[2]-=- score normalization is defined as: ˆsij = 1 (1 + e −t ) , t = sij − µi σi (10) where ˆsij is a normalized similarity score, sij is a raw similarity score between image/subject i and image/subject j, ...

Citation Context

...he feature vector consisted of 15 static Mel-PLP coefficients, 15 delta coefficients and the delta energy. Static energy or acceleration features were not extracted as it has been previously reported =-=[8]-=- that they contain no speaker-discriminative information. Lastly, Cepstral Mean Subtraction was performed on each utterance to introduce robustness to stationary channel noise. Cepstral feature warpin...

Citation Context

...es have been used to decide if a particular frame in a test song contain vocals or not. To our knowledge, these world models have not been evaluated for the use in score normalization as described in =-=[4]-=-. Also, ways to optimize the vocal/nonvocal decision based on world models have not been evaluated. Furthermore, we would like to examine a straightforward approach to eliminate some influences of the...