A study on the speech analysis and recognition for Korean children

This paper describes a speaker‐independent Korean ten‐digit recognizer for males and females from 10 to 13 years old, who have a serious variation of speech characteristics. DHMMs with time normalized ten‐dimensional LPC, CEPSTRUM, LSF coefficients, and ten‐band energies in frequency domain as the feature vector are used. The pitch and formant informations are extracted to analyze speech characteristics. Through experiments, DHMM recognizer shows the performance of recognition rates, 67.1% (LPC vectors), 80.8% (CEPSTRUM vectors), 66.7% (band energies in frequency domain), and 83.8% (LSF vectors).