2002 Fiscal Year Final Research Report Summary
Asynchronous-Transition Hidden Markov Model with State-Tying across Time for Automatic Speech Recognition
| Project/Area Number |
12680375
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Japan Advanced Institute of Science and Technology |
|---|
Principal Investigator |
SHIMODAIRA Hiroshi JAIST, School of Information, Science, Associate Professor, 情報科学研究科, 助教授 (30206239)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
NAKAI Mitsuru JAIST, School of Information Science, Research Associate, 情報科学研究科, 助手 (60283149)
SAGAYAMA Shigeki The University of Tokyo, Graduate School of Information Science and Technology, Professor, 大学院・情報理工学系研究科, 教授 (00303321)
|
|---|
| Project Period (FY) |
2000 – 2002
|
| Keywords | hidden Markov model / HMM / asynchronous-transition / AT-HMM |
|---|
| Research Abstract |
This project aimed to improve acoustic models for speech recognition systems. The state-of-the-art hidden Markov model (HMM) based acoustic models usually treat the acoustic features as a chain of stationary signal sources. The observed values of these features are represented by vectors. We assumed that they might be better modeled by individual vector components. We discussed two methods based on this assumption
In the first method, wearied to model asynchronous changes of individual acoustic vector components. Conventional HMM implicitly assumes that individual components change their statistical properties simultaneously. This assumption might be not true. Temporally changing patterns of individual acoustic components do not necessarily synchronize with beach other. We proposed a new HMM that allowed asynchronous state transitions between individual vector components. We demonstrated that this new HMM outperformed the conventional HMM in speaker-dependent speech recognition task
In the second method, we tried to model phoneme context dependency of individual acoustic vector components. Conventional parameter tying techniques provide a common tying structure for all vector components, no matter how different is their individual components complexity and phoneme context dependency. In this discussion, we proposed a new parameter tying technique that allowed to have distinct tying structures for each component. Our experimental results showed that proposed HMM with feature-depended tying worked better than conventional HMM with a common tying
|
