PURPOSE: To reduce deterioration in speech recognition rate due to environmental changes by finding and matching the difference vector from an input vector, generated while a mean vector is updated at the time of recognition, with a standard pattern.
CONSTITUTION: When a speech section is decided and the mean vector is calculated, a mean vector arithmetic part 408 performs the calculation and stores the found mean vector in a mean vector register 409. Then a difference vector arithmetic part 410 finds the difference vector by subtracting the mean vector in the mean vector register from the input vector of an input speech and stores the difference vector in a difference vector register 411. Then when the speech recognition is indicated, a series of difference vectors of the input signal of the speech is read out of the difference vector register 411 and matched by a pattern matching part 414 by using word difference standard patterns 413 to calculate the differences between the respective word difference standard patterns and the input pattern, and on the basis of the result, the recognition output is sent out through a recognition decision part 415.
AKAHA TOSHIO
HAMAGUCHI SEIJI
