Training of speech reference patterns to situation dependent pronunciation variants

摘要

Investigations have shown that the results of speech recognition devices depend not only on the acoustic environment (background noises) but also on the psychological condition of the speaker (stress). According to the invention, a continuous training of stored speech reference patterns to pronunciation variants occurs in that, for each correctly recognised word, a new speech reference pattern is calculated from the stored and the newly recognised pattern by means of weighted averaging and is stored instead of the previous pattern.