摘要 |
<p>Process for adaptation of a hidden Markov sound model in a speech recognition system. This invention concerns a process for adapting a generally available code book (CB) for special applications with a speech recognition system of the hidden Markov sound model. These applications are defined by a lexicon (LEX) changed by the user. The adaption (ADAP) is done during operation and occurs by means of a displacement of the stored midpoint vector of the probability density distributions of hidden Markov models, in the direction of a known feature vector of sound expressions and in relationship to the hidden Markov models specially used. In comparison to current practices, the invention has the advantage that it is done on-line and that it has a very high recognition rate with little computational expenditure. In addition, the training expenditure for special sound models for corresponding applications is avoided. By using special hidden Markov models from multilingual phonemes, in which sound similarities across various languages are used, automatic adaptation to foreign languages can follow. Both language-specific and language-dependent characteristics are taken into account in this method for acoustic phonetic modelling to determine the probability densities for different hidden Markov sound models in different languages.</p> |