A method for determining a dead space to the representation of a plurality of training speakers

摘要

Disclosed is a process for determining a dead space to the representation of a plurality of training speakers, in which initially speaker-dependent model - sets for the individual training with the use of training speech data of the individual training be formed, whereby the models (sd) of a model - sets in each case by a plurality of model parameters will be described. For each speaker is then a combined model in many-dimensional vector space by chaining a of the model parameters of the models of the individual training set in each case to form a coherent. Subsequently, a transformation with a reduction in the dimension of the model space base vectors of space - (e φ e φ) is carried out. In order to achieve an unambiguous assignment of the model parameters in the super vectors to ensure, in this case, firstly a common speaker-independent model - kit for the said training speakers have been developed and this model - set to the individual training to the development of the speaker-dependent model - sets adapted. The assignment of the model parameters of the models (si) of the speech model - sets to the model parameters of the models (sd) of the speaker-dependent model - sets in this case is detected and the concatenation of the model parameters of the individual model - sets to the super vectors is carried out with consideration of this assignment.