Presentation | 2007/12/13 Speaker verification using multi-stream HMMs with dimensionally weighted feature vectors Shinya KOJIMA, Koji IWANO, Sadaoki FURUI, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | This paper describes a noise-robust speaker verification method using multi-stream HMMs. In this framework, each dimension of acoustic feature vectors is treated as an individual stream and then each stream is automatically weighted by a Linear Discriminant Analysis (LDA) based technique so as to emphasize noise-robust dimensions. This paper proposes two types of weight estimation using the LDA-based technique : unsupervised estimation and supervised noise-independent-weight estimation. In the latter estimation, the LDA is applied to development data contaminated with various noises different from the noise in testing data. Experiments were conducted using Japanese four-connected-digit utterances in various kinds of noise and SNR conditions. Two kinds of acoustic features, cepstrum domain features (MFCC) and spectrum domain features (SPEC),were used for the experiments. For both features, experimental results show that the unsupervised estimation reduces the equal error rates (EERs) in relatively high-SNR (15-20dB) conditions and the noise-independent-weight estimation improves the noise robustness in all noise conditions. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | speaker verification / noise robustness / multi-stream HMM / linear discriminant analysis (LDA) / spectral domain feature |
Paper # | NLC2007-40,SP2007-103 |
Date of Issue |
Conference Information | |
Committee | NLC |
---|---|
Conference Date | 2007/12/13(1days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | |
Vice Chair | |
Secretary | |
Assistant |
Paper Information | |
Registration To | Natural Language Understanding and Models of Communication (NLC) |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Speaker verification using multi-stream HMMs with dimensionally weighted feature vectors |
Sub Title (in English) | |
Keyword(1) | speaker verification |
Keyword(2) | noise robustness |
Keyword(3) | multi-stream HMM |
Keyword(4) | linear discriminant analysis (LDA) |
Keyword(5) | spectral domain feature |
1st Author's Name | Shinya KOJIMA |
1st Author's Affiliation | Department of Computer Science, Tokyo Institute of Technology() |
2nd Author's Name | Koji IWANO |
2nd Author's Affiliation | Department of Computer Science, Tokyo Institute of Technology |
3rd Author's Name | Sadaoki FURUI |
3rd Author's Affiliation | Department of Computer Science, Tokyo Institute of Technology |
Date | 2007/12/13 |
Paper # | NLC2007-40,SP2007-103 |
Volume (vol) | vol.107 |
Number (no) | 405 |
Page | pp.pp.- |
#Pages | 5 |
Date of Issue |