METHOD FOR COMPRESSION OF THE COMPOSITION OF LEXICON WITH THE LANGUAGE MODEL IN A SYSTEM OF RECOGNITION OF TALK FOR MOBILE DEVICES

摘要

The present patent discloses a method of compressing the composition of a lexical with a language model in the framework of a decoding system of a speech recognition system. THIS METHOD CAN BE PERFORMED ON A LARGE VARIETY OF DEVICES, ALLOWING THE EXECUTION OF MEDIUM-SIZED RECOGNITION TASKS ON DEVICES WITH COMPUTATIONAL RESOURCES, AND FACILITATING THE PARALLELIZATION OF TASKS WITH VERY LONG VOCABULARS IN MACHINES OF HIGHER COMPUTATIONAL RESOURCES. THE METHOD DESCRIBED HEREIN ESSENTIALLY UNDERSTANDS TWO MAIN POINTS: COMPRESSION IN TIME OF EXECUTION OF THE COMPOSITION OF LEXICON WITH THE LANGUAGE MODEL USING LINEAR PATHS AND A VARIABLE SIZE CODING OF THE FIELDS OF THE MODEL ARCHES, AND THE DRAWING OF A DATA STRUCTURE ( CACHE) TO FACILITATE ACCESS TO THE COMPRESSED TRANSDUCER, CREATED IN ORDER TO EXPLORE THE AUTOMATED PARTICULARS (LINEAR PATHS) USED IN THE PREVIOUS STEP.