| Project/Area Number |
16500092
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Waseda University (2005-2006)
The Institute of Statistical Mathematics (2004) |
|---|
Principal Investigator |
TANABE Kunio Waseda University, Faculty of Science and Engineering, Professor (50000203)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
MATSUI Tomoko Waseda University, Institute of Statistical Mathematics, Dept.of Modelling, Associate professor (10370090)
|
|---|
| Project Period (FY) |
2004 – 2006
|
| Project Status |
Completed (Fiscal Year 2006)
|
| Budget Amount *help |
¥3,600,000 (Direct Cost: ¥3,600,000)
Fiscal Year 2006: ¥1,200,000 (Direct Cost: ¥1,200,000)
Fiscal Year 2005: ¥1,400,000 (Direct Cost: ¥1,400,000)
Fiscal Year 2004: ¥1,000,000 (Direct Cost: ¥1,000,000)
|
|---|
| Keywords | Learning Machine / Inductive Reasoning / Computational Algorithm / Penalized Likelihood / Logistic Regression / Discrimination / Speech Recognition / Speaker-identification / Universal Induction Machine / Learning Machine / Automatic Discrimination / Probabilistic Prediction / Multimodal Data / Speach Recognition / Speaker Identification / Sound Source Detection / 罰金付き / 異種混合データ / 統合的判別 / カーネルマシン |
|---|
| Research Abstract |
By employing the dual Penalized Logistic Regression Machines (dPLRM) which were proposed by K. Tanabe in 2001, a speaker identification method which does not require feature extraction pre-processing was developed. The induction machine dPLRM can discover implicitly hidden speaker characteristics relevant to discrimination of possible speakers dpending only on a set of training data without resorting to available information of voice characteristic developed in the past 30 in the speech recognition field. Our text-independent speaker identification experiments with training data uttered by male speakers in three different sessions show that the proposed method is competitive with the conventionalGaussian mixture model(GMM) -based method with 26-dimentional Mel-frequency cepstrum(MFCC) feature even though our method handle directly coarse data of 256-dimensional log-power spectrum. It was also shown that our method outperforms the GMM-based method especially as the amount of training data becomes smaller.
The universal induction machine dPLRM requires heavy numerical computation. We have developed a computational algorithm for huge-scale linear equation solver to be employed in dPLRM. We also proved the convergence of the Rump's method for inverting a vetry ill-conditioned matrices.
|
