휴대 전화기 상에서 Name Dialing을 위한 화자독립 실시간 음성인식
Speaker-Independent Real-Time Speech Recognition for Korean Name Dialing on Mobile Phone
v, 38 p.
휴대전화 Name Dialing 화자독립 실시간 음성인식;
- 원문 URL
A speaker independent speech recognition system for Korean name dialing on mobile phone is proposed in this paper. This system based on CDHMM overcomes the limitations of DTW based models for speaker dependent speech recognition for name dialing while meeting the requirements of mobile phone on limited resources regarding memory size and CPU clock speed. In order to meet the requirements, several conventional algorithms such as TSVQ and PBS are used and a new method to reduce the pdf computation time in GMMA method is proposed. The newly proposed method for reducing the pdf computation in GMMA method finds the input features that can be considered as minor ones in deciding maximum Gaussian mixture density for GMMA method. The proposed method then removes the insignificant input features from calculating observation probability in GMMA method. The speed of observation probability calculating of CDHMM is improved by using the proposed new algorithm: improvement of instruction cycles is about 16.8% with a degradation of the recognition accuracy by less than 0.2%. Experiments on the Korean name dialing under real mobile phone environment inside of automobiles show that the proposed system achieves about 92% of recognition accuracy with about 2.4 seconds of response time.