Procedure for creating reference segments describing speech modules from signals training voice pronounced for systems speech recognition, comprising the following steps: - segmentation of the speech signal training speech modules corresponding to a predetermined transcript , - analysis of the training signal in a temporal grid predetermined certain time windows for at least one characteristic vector for each time window, such that training models each containing eigenvectors case in temporal succession are configured the speech signal training, - determination of an intermediate temporary structure that is an intermediate manifestation of the length variation and the time sequence of the characteristics of the speech modules, for each speech modules with a sequence time windows, - association of vecto characteristic by a nonlinear time windows of the speech modules and storage of the characteristic vectors each associated to a speech module in the predetermined sequence by the time windows as a reference segment temporary reproduction res.
展开▼