Martin Holzapfel

Determining an adequate representative sound using two quality criteria, from sound models chosen from a structure including a set of sound models

A method determines a representative sound on the basis of a structure which includes a set of sound models. Each sound model has at least one representative for the modeled sound. In the structure, a first sound model, matching with regard to a first quality criterion, is determined from the set of sound models. At least one second sound model is determined from the set of sound models dependent on a characteristic state criterion of the structure. At least some of the representatives of the first sound model and of the at least one second sound model are assessed in addition to the first quality criterion with regard to a second quality criterion. The at least one representative which has an adequate overall quality criterion with regard to the first and second quality criteria is determined as a representative sound from the representatives of the first sound model and the at least one second sound model.

Optimization of a neural network for speaker and task dependent F/sub 0/-generation

The generation of a pleasant pitch contour is an important issue for the naturalness of each TTS system. Until now the results are far from being satisfactory. We present a speaker and task specific approach realized by a neural network. Personal and task specific characteristics are maintained and the demand of generalization decreases. Therefore the results in application can significantly be improved. Using an optimized network structure global and well localized patterns can be covered and trained simultaneously within one network. Correlation analysis of the data base versus the sensitivity of the trained network validates the importance of distinctive parameters in training. Based on this comparison we discuss the generalization properties of the NN trained speaker and task dependency. Finally a variation of the context range helps to find an optimized tuning of the input parameter set.

Automatic learning of numeral grammars for multi-lingual speech synthesizers

Presented is a trainable data-driven method of deriving numerals from number strings. The concept is based on learning a graph model for numeral grammars and a graph search which is capable of extracting numeral words from the grammar according to given number strings. Because this method separates code and data it is universal and applicable to every language. No a priori knowledge about formal descriptions of numeral grammars is required. Due to the underlying graph concept, the algorithm is able to automatically generate grammars for number to numeral translations with a minimal effort. The only required input information is a set of pairs of number strings and appropriate numerals. This method is useful for an easy implementation of knowledge bases for further languages in the framework of a multi-lingual speech synthesis system.