DEVICE FOR GENERATING A REFERENCE PATTERN WITH A CONTINUOUS PROBABILITY DENSITY FUNCTION DERIVED FROM FEATURE CODE OCCURRENCE PROBABILITY DISTRIBUTION

摘要

In a system for recognizing a time sequence offeature vectors of a speech signal representative of anunknown utterance as one of a plurality of referencepatterns, a generator (11) for generating the referencepatterns has a converter (15) for converting a pluralityof time sequences of feature vectors of an input patternof a speech signal with variances to a plurality of timesequences of feature codes with reference to codevectors (14) which are previously prepared by the knownclustering. A first pattern former (16) generates astate transition probability distribution and anoccurrence probability distribution of feature codes foreach state in a state transition network. A functiongenerator (17) calculates parameters of continuousGaussian density function from the code vectors and theoccurrence probability distribution to produce thecontinuous Gaussian density function approximating theoccurrence probability distribution. A second patternformer (18) produces a reference pattern defined by thestate transition probability distribution and thecontinuous Gaussian density function. For a pluralityof different training words, a plurality of referencepatterns are generated and are memorized in thereference pattern generator.