ISCA Archive SPKD 2008
ISCA Archive SPKD 2008

Feature selection algorithms for the creation of multistream speech recognizers

Yotaro Kubo, Shigeki Okawa, Akira Kurematsu, Katsuhiko Shirai

In this paper, we present a method to split a feature stream into multiple feature streams.

The efficiency of ensemble classifiers for speech recognition is confirmed by several experiments. The conventional methods for constructing multiple classifiers are done by splitting the feature stream by type of features or subbands where the features are associated. The splitting approach is well suited for obtaining high-dimensional features because it naturally leads to dimension reduction of features. In order to take advantage of ensemble classifiers, each classifier should compensate for the errors due to the other classifiers. Because every streams depends on phonetic information in clean environments, the difference on noise robustness can be measured by using independency. Therefore, each classifier should be independent from others in noisy environments. We proposed a method to split a feature stream using stream independency criteria in order to constructing independent classifiers. We evaluated several stream splitting methods and compare word error rate by conducting continuous digit recognition experiments on noisy speech. Our method can reduce 30.9% of the word error when compared with the single classifier method, while it reduces 3.2% of the word error when compared with conventional multistream approach.


Cite as: Kubo, Y., Okawa, S., Kurematsu, A., Shirai, K. (2008) Feature selection algorithms for the creation of multistream speech recognizers. Proc. ISCA ITRW on Speech Analysis and Processing for Knowledge Discovery, paper 049

@inproceedings{kubo08_spkd,
  author={Yotaro Kubo and Shigeki Okawa and Akira Kurematsu and Katsuhiko Shirai},
  title={{Feature selection algorithms for the creation of multistream speech recognizers}},
  year=2008,
  booktitle={Proc. ISCA ITRW on Speech Analysis and Processing for Knowledge Discovery},
  pages={paper 049}
}