|
Due to the low Mandarin consonant recognition rate , in this
thesis , a new recognition network named Dynamic
Programming Bayesian Neural Network is proposed to improve
the consonant recognition rate. In this network , we combine
the dynamic model and Bayesian network which will overcome
the problem of the dynamic feature of speech signal and
tolerate spectral pattern variation. Furthermore , since the
feature parameters selection is very important for consonant,
we use cepstrum coefficient , delta cepstrum coefficient and
time domain features such as energy and zero-crossing rate to
represent the feature of consonant . In this paper, a
hierarchical recognition scheme is used to reduce the matching
times to accelerate recognition speed. Consequently , the
number of candidates can be reduced from 408 to 64 , i.e. , 4
lexicals , 38 vowels ,22 consonants. We use a segmentation
algorithm to segment the connected speech into syllables , and
each syllables is then segmented into vowel and consonant
parts . For experimental evaluation,we build a database
composed of two males , each one utters 408 Mandarin syllables
seven times . Four of them are used as training patterns and
three for testing patterns . After a series of experiments ,
the top one recognition rate of 87.59% is achieved and the
top five recognition rate is 98.51% in average .
|