A Design of Bilingual Character and Speech Recognition System for Japanese and Korean

• Main Authors: Li-Chung Hsu, 許力中
• Other Authors: Chih-Chien Chen
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/93449728685407804757

碩士 === 國立中山大學 === 電機工程學系研究所 === 102 === Traveling becomes more and more popular in the recent years. According to the statistics of Tourism Bureau, Ministry of Transportation and Communications, ROC, the number of tourists abroad grew at a steady pace during the past few years, especially in the neighboring Japan and Korea. The tourist population from Taiwan to Japan and Korea are ranked number three and number four, just fewer than that of Mainland China and Hong Kong. Chinese culture has been influencing language, custom and tradition for both countries since 200 B.C. Chinese characters were used as their ancient writing systems. Commercial signboards in Kanji can be found today all over the Japanese streets. And more than 70% of the Korean vocabularies can be pronounced from ancient Chinese. A unique Chinese cultural circle in Asia is forming under the realm of the three languages. In Taiwan, Japanese and Korean waves have significant impacts from the show business to the life styles of common people. Drama and fashion are the favorites. Hence, it is our objective to establish a character and speech recognition system for Japanese and Korean to learn the languages, to experience the cultures and to widen our perspectives as well. In this thesis, a bilingual character and speech recognition system for Japanese and Korean is designed and implemented. Two-dimensional Fourier transform and Karhunen-Loeve transform are used to extract the character features of 152 Japanese Kana classes, 2,427 Japanese Kanji classes and 984 Korean syllable patterns. Cosine similarity of the features and literal structure of the two languages are then applied to find the final answer. Under the 2.3 GHz Intel Core i5 PC and Windows 7 operating system environment, correct character recognition rates of 98.86% and 97.56% can be reached respectively for the 5,000 Japanese word and 5,000 Korean word databases. Mel-frequency cepstral coefficients and linear predicted cepstral coefficients are utilized for the speech feature extraction of the selected 207 Japanese and 712 Korean common syllables. Hidden Markov model and phonotactics are then employed to obtain the ultimate solution. Under the 2.5 GHz Intel Core2 Quad PC and Ubuntu 12.04 operating system environment, correct speech recognition rates of 95.83% and 95.62% can be attained respectively for the 4,485 Japanese word and 4,546 Korean word databases.