| Presentation | 2004/12/13 Robustness of acoustic model topology determined by VBEC for different speech data sets Shinji WATANABE, Atsushi NAKAMURA, |
|---|---|
| PDF Download Page | PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | A lack of robustness with acoustic modeling often degrades the performance of spontaneous speech recognition and understanding. One reason for this shortcoming is that the Maximum Likelihood (ML) approach based on model parameter estimation has a poor generalization ability. This makes it important to improve the generalization ability of robust training of models including HMM and future techniques beyond HMM. The Bayesian approach is based on posterior distribution estimation, and has a better generalization ability than the ML approach due to the marginalization effect of model parameters. Variational Bayesian Estimation and Clustering for speech recognition (VBEC) is a total Bayesian framework in the sense that all speech recognition procedures are based on posterior distribution estimation within the Variational Bayes method, which includes the Bayesian advantage of highly generalized model training. In addition, a VBEC specification of the posterior distribution estimation enables automatic determination of an acoustic model topology without heuristics, by regarding model complexity as a probabilistic variable, and by selecting the appropriate model that scores the maximum probability value. In this paper, we describe experiments for different speaking-style (isolated word, continuous speech and spontaneous lecture speech) and language sets (Japanese and English) of training data, and show the effectiveness of VBEC, which automatically determines the model topology robustly according to the speech types of the training data. We also examine the robustness of the determined models for a mismatched condition between training and test data tasks. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Speech recognition / VBEC / Automatic determination of acoustic model topology / Robustness for speaking style, / language and mismatched condition between training and test data |
| Paper # | NLC2004-50,SP2004-90 |
| Date of Issue |
| Conference Information | |
| Committee | NLC |
|---|---|
| Conference Date | 2004/12/13(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 | ENG |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Robustness of acoustic model topology determined by VBEC for different speech data sets |
| Sub Title (in English) | |
| Keyword(1) | Speech recognition |
| Keyword(2) | VBEC |
| Keyword(3) | Automatic determination of acoustic model topology |
| Keyword(4) | Robustness for speaking style, |
| Keyword(5) | language and mismatched condition between training and test data |
| 1st Author's Name | Shinji WATANABE |
| 1st Author's Affiliation | Nippon Telegraph and Telephone Corporation, NTT Communication Science Laboratories() |
| 2nd Author's Name | Atsushi NAKAMURA |
| 2nd Author's Affiliation | Nippon Telegraph and Telephone Corporation, NTT Communication Science Laboratories |
| Date | 2004/12/13 |
| Paper # | NLC2004-50,SP2004-90 |
| Volume (vol) | vol.104 |
| Number (no) | 538 |
| Page | pp.pp.- |
| #Pages | 6 |
| Date of Issue | |