LPC 계수를 이용한 음성인식 시뮬레이션

소개글

LPC 계수를 이용한 음성인식 시뮬레이션에 대한 보고서 자료입니다.

목차

1.object

2.theory & Algorithm

3.result analysis

4.afternote

5.reference

본문내용

p the decision range by calculating average and made decision range value 1. Therefore in above figure average of my_voices is 20.7263 so decision range si 19.7263~21.7263. Other_voice(my frend's voice) is over range so my program distinguish it's not my voice and show '0' As we can see, all of my_voices(Final_Result) are in the accurate range(19.7263~21.7263) so if Other_voice is one of my_voices, my programe distinguish it's my voice and show '1'.
Figure 8. Equation of Cost function
Cost function is like above. In my simulation with my friend's voice, due to shortage of samples, so C=0. But If I have more samples I can calculate C more accurately.
There are some problems in my program. In this simulation I set up range is 1. But If I say too low voice or too high voice, my voice's pitch will be too high or too low. It will be over range and my system will distinguish it's not my voice. To solve this problem, I can try to say again like when I recorded or I can increase range value to 2 or 3 ... But If I do like the latter, there will be more error because I expand decision range. So when I say the password to open my vault, I have to say carefully.
Furthermore, in the simulation pitch period of all my voices are similar. But when I recorded one or two my voice, if I say too high voice or too low voice it affects the average of 10 of my voices - decision average goes up or goes down. It will be serious problem. To solve this problem, I can record carefully to say in similar tone or when I calculate average of 10 of my voices, I can except the highest pitch period and lowest pitch period - calculate average out of 8 voices. If I used GMM to distinguish voice, it will be better performance and it doesn't make serius problem like this but I couldn't make. In next project I will study much harder and try to make perfect solution.
: Afternote
Thanks for reading very very long report. Through this project, I could study a lot of speech recognition and learn and get interested in speech signal process. It will be helpful to study in graduate school and I learn make constant efforts to the impossible like problem.
: Reference
[1] J. W. Picone, "Signal modeling techniques in speech recognition," Proc. IEEE, vol. 8, no. 9, pp. 1215-1247, Sept. 1993.
[2] D. O'Shaughnessy, Speech Communication: Human and Machine, 2nd ed. Addison-Wesley.
[3] J. B. Allen, "Cochlear modeling," IEEE ASSP Magazine, vol. 3, no. 3, pp. 3-29, Sept. 1985.
[4] S. B. Davis and P. Mermelstein, "Comparison of parametric representations of monosyllabic word recognition in continuously spoken sentences," IEEE Trans. ASSP, vo. 28, no. 4, pp. 357-366, Aug. 1980.
[5] H. Hermansky, et al., "RASTA-PLP speech analysis technique," ICASSP'92.
[6] J.-C. Junqua and J.-P. Haton, Robustness in automatic speech recognition, Kluwer Academic Publishers, 1996.
[7] 음성인식 , 한양대학교 출판부
[8] 한학용 , 패턴인식개론 , 한빛미디어
[9] 한진수 , 음성 신호 처리 , 오성미디어
[10] Google search by "Speech recognition", "LPC", "GMM", "MFCC" ...

키워드

lpc,   음성인식,   음성검출,   암호화