Answers are grounded in this paper and cite their sources.
IntroductionIt has been demonstrated throughout the literature that there are recognition gains to be had by clustering available training data according to speaker sex, channel, noise conditions or speaking rate.Essentially, these results demonstrate that models conditioned according to categories of variability are more effective than unconditioned models.Recently, there have been a number of papers extending this concept in systematic ways.Padmanabhan et al.move from two gender clusters to many speaker clusters in 1996 [1].In 1998, Kuhn et al. introduced eigenvoices, a method that provides both a novel mechanism of clustering and an efficient adaptation [2).Both of these works promise rapid adaptation due to the conditioning of the pre-trained cluster mod els.Looking at the matter another way, the adaptation mechanism has been partially effected in the training procedure.Given that training corpora are continually expanding, it makes sense to.exploit their richness as much as possible.Byproviding tor macro-adaptation in this manner, the on-line adaptation procedure now reduces to micro-adaptation.As a result, it now requires much less adaptation data to achieve successfully adapted models.We propose an eigenvoice-like architecture in which adapted models are constructed as a linear combination of pre-trained cluster models, deviating from the eigenvoice approach in four important ways:1. phonetic class-based cluster models, 2. data-driven cluster model definition, 3. non-linear programming for cluster coefficient calculation, & 4. cluster coef cient export to high-definition models.Recently, eigenvoices and speaker clustering have taken turns toward class-dependent modeling.Kenny, Boulianne and Dumouchel consider cluster modeling at the Gaussian level [3].Jiang and Huang observe gains by adding subword dependence to speaker clustering [4].Considering that the modeling-unit dependence of traditional adaptation methods has been considered in great detail, this research trajectory is not surprising.Unfortunately, the smaller the granularity of unit-dependence.
No takes yet. Share an insight, caveat, or question.
S.D. Peters (2024) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: