2001 Fiscal Year Final Research Report Summary
Computer-Assisted Pronunciation Learning System using Speech Recognition Techniaues
| Project/Area Number |
11558037
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (B)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 展開研究 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | KYOTO UNIVERSITY |
|---|
Principal Investigator |
KAWAHARA Tatsuya Kyoto University, Graduate School of Informatics, Associate Professor, 情報学研究科, 助教授 (00234104)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
KATAGIRI Shigeru NTT Communication Science Laboratories, Executive Manager, コミュニケーション科学基礎研究所, 研究部長
DOSHITA Shuji Ryukoku University, Faculty of Science and Technology, Professor, 理工学部, 教授 (00025925)
DANTSUJI Masatake Kyoto University, Center for Information and Multimedia Studies, Professor, 総合情報メディアセンター, 教授 (10188469)
SHIMIZU Masaaki Kyoto University, Center for Information and Multimedia Studies, Assistant Professor, 総合情報メディアセンター, 助手 (10314262)
OKUNO Hiroshi Kyoto University, Graduate School of Informatics, Professor, 情報学研究科, 教授 (60318201)
|
|---|
| Project Period (FY) |
1999 – 2001
|
| Keywords | speech processing / language learning / CALL / speech recognition / phonology / prosody |
|---|
| Research Abstract |
A Computer-Assisted Language Learning (CALL) system focusing pronunciation training is studied for English learning by Japanese students.
First, we model typical English pronunciation errors of Japanese learners and design a system that detects pronunciation errors and generates -effective instruction utilizing speech recognition technologies. For a given training text, a network of error candidates is generated for speech recognition to align the utterance and detect errors. Then, a segment-input pair-wise classifier is applied forverification. This method realizes reliable errordetectionandeffective instruction based on articulatory information.
Then, we develop a computer-assisted English prosody learning system. Learners' pronunciation is evaluated by automatic detection of sentence stressed syllables and foot durations. Syllable HMMs are categorized based on error patterns of stress. We also propose a method of multi-stage discrimination that reflects native speakers' perception. Furthermore, foot templates are constructed from native speech database in order to evaluate stress-timing.
Finally, we study to estimate non-native speakers' intelligibility and to determine which pronunciation errors affect intelligibility the most. A preliminary study showed that error rates computed by a speech recognition-based system can be used to characterize intelligibility. We use the error rate distributions to assess the student's intelligibility and compute a priority function to find which areas of study are most likely to improve the intelligibility.
|
