This paper proposes the application of discriminative training techniques based on the generalized probabilistic descent (GPD) approach to stochastic Markov graphs (SMGs), a generalization of mixture-state hidden Markov models (HMMs), describing the constraints in the acoustic structure of speech as a graph consisting of nodes, each containing a base function, and a transition network between the nodes. State-specific weights modeling the classification relevance of the corresponding states and a transition weight representing the ratio between transitions and emissions are trained in addition to the standard parameters of the models. The experiments show, that discriminatively trained SMGs outperform discriminatively trained mixture-state HMMs with approximately the same number of parameters.

We present an algorithm to identify potential functional elements like protein binding sites in DNA sequences, solely from nucleotide sequence data. Prerequisites are a set of at least seven not closely related sequences with a common biological function which is correlated to one or more unknown sequence elements present in most but not necessarily all of the sequences. The algorithm is based on a search for n-tuples which occur at least in a minimum percentage of the sequences with no or one mismatch, which may be at any position of the tuple. In contrast to functional tuples, random tuples show no preferred pattern of mismatch locations within the tuple nor is the conservation extended beyond the tuple. Both features of functional tuples are used to eliminate random tuples. Selection is carried out by maximization of the information content first for the n-tuple, then for a region containing the tuple and finally for the complete binding site. Further matches are found in an additional selection step, using the ConsInd method previously described. The algorithm is capable of identifying and delimiting elements (e.g. protein binding sites) represented by single short cores (e.g. TATA box) in sets of unaligned sequences of about 500 nucleotides using no information other than the nucleotide sequences. Furthermore, we show its ability to identify multiple elements in a set of complete LTR sequences (more than 600 nucleotides per sequence).

This paper proposes a new modeling of the structure of speech units as a graph consisting of base functions and a transition network. A cluster algorithm taking into account the actual temporal context of the feature vectors is used to generate the base functions, which are approximated by normal distributions. The subsequent Viterbi-based maximum-likelihood training procedure establishes the transition network and adjusts the transition probabilities. The emerging graphs for the speech units are a structure of branching and recombining trajectory segments describing statistical dependencies in the feature vector sequence within the speech units as well as in the transition regions between them. A speaker-independent evaluation shows the superiority of the proposed modeling compared to mixture-state HMMs, even for an equal number of model parameters.

In this paper neural networks for Nonlinear Discriminant Analysis in continuous speech recognition are presented.Multilayer Perceptrons are used to estimate aposteriori probabilities for Hidden-Markov Model states, which are the optimal discriminant features for the separation of the HMM states.The a-posteriori probabilities are transformed by a principal component analysis to calculate the new features for semicontinuous HMMs, which a r e trained by the known Maximum-Likelihood training.The nonlinear discriminant transformation is used in speakerindependent phoneme recognition experiments and compared to the standard Linear Discriminant Analysis technique.