A  Study to  Consonant/Vowel  Phonetic  Segmentation 
of Korean Isolated Word based on a Rule-based System 
for the Phenomenon of Korean vocalization

     In speech recognition, the method of phoneme is required to compensate for much database volume and quantity of calculation needed for getting a recognition method such as the method of syllables or words.  But, in the case of Korean speech analysis, Korean phonetic changes are illuminated by the method of phonetic value than the method of phoneme. 
 
 

     In this paper, we have been implemented and verified a rule-based system for the characteristic phenomenon of Korean vocalization that can be used as a supporting tool of Korean speech processing.  And consonant/vowel segmentation of phonetic value for Korean isolated word is perfomed as the pre-processing unit for the recognition method of phonetic value where consonant/vowels segmentation of the phonetic value is carried out based on a rule-based system.
 
 
 
 
 
 
 
 

    First, as a pre-processing step of consonant/vowel segmentation of phonetic value, the speech signal of Korean isolated word is analyzed by wavelet transform that has similar to features with human hearing model. After the multi-resolution decomposition of speech signal using wavelet transform that has to discrete Daubechies' wavelet basis, the speech range detection of Korean isolated word is carried out using energy of wavelet coefficient.
 

     And as a using standard deviation of wavelet bandwidth channel that shows a sudden change, preliminary consonant/vowel segmentation border points of phonetic value are extracted. The border points extracted are verified that include the real phonetic value border points acquired by manual. 
 
 
 
 
 
 
 
 
 

     Finally, the nonuniform blocks are extracted in the range of preliminary border points as a result. In the non-uniform block, the extracted essential feature parameters is separated to each five small overlapped uniform block (state each five small overlapped uniform block) in the non-uniform block: energy, zero-crossing, level crossing, 20th order LPC (linear predictive coefficient) cepstrum, energy ratio of FFT (fast Fourier transform) frequency band, 9th order wavelet coefficient etc. 
 

     The extracted feature parameters in the each five small overlapped non-uniform block divided as a several group. And the each extracted feature parameter group of the small uniform block accomplishes each single codebook using vector quantization featuring a standard K-means algorithm to decrease the pattern domains. Using K-means VQ algorithm, the vector quantized multi-codebook to be formed by a single code book of each feature parameter group is recreated a single codebook which contain a various signal feature. 
 
 
 
 
 
 
 
 
 
 

      Finally, codeword index sequence of quantized vectors in the final single codebook are used as an input parameter for DHMM(discrete hidden Markov model). Extracted a feature parameter vector from the analysis of non-uniform block and received to an input DHMM from index of similarity codeword of final single codebook by forward-backward algorithm. Observation sequence of imaginary an input speech pattern of non-uniform block provided to consonant/vowel DHMM model by Viterbi algorithm. And, observation sequence of final codebook index sequence feature vectors and current DHMM parameter re-estimated to consonant/vowel DHMM parameter from Baum-Welch algorithm. Finally, a consonant/vowel segmentation of phonetic value that is based on statistical possibility is accomplished.