Chapter 2 starts with an introduction to the copula model and dierent multivariate copula
functions. Then, we propose two computationally simple approaches to built multivariate
copula functions in Chapter 3. Finally, we evaluate these approaches in dierent density
estimation and classication tasks. In chapter 4, we propose two simple methods to address
the mismatch between training and testing conditions, which is a major bottleneck in
everyday applications of ASR systems. First, we introduce a simple copula-based feature
enhancement method independent of the ASR backend. Later in this chapter, we show
that this enhancement can be further optimized by integrating it into the acoustic model.
For evaluation, we use large vocabulary speech recognition tasks on Aurora 4 and Chime4.
Chapter 5 gives the conclusions and future work.
We evaluate the performance of both of the proposed methods on several density estimation tasks from the UCI Repository as well as our corpus of Monkey vocalizations,
recorded at the Oregon National Primate Research Center. We ﬁnd that our methods represent the data consistently better than Gaussian mixture models with equivalent number
of parameters. We also evaluate our proposed methods on building generative classiﬁers
for a number of classiﬁcation tasks from the UCI Repository. We ﬁnd that these generative models perform as well or better than discriminative classiﬁers such as a Support
Vector Machine (SVM).