Detail publikace

Data adaptation for Hidden Markov Model in speech recognition

Originální název

Data adaptation for Hidden Markov Model in speech recognition

Anglický název

Data adaptation for Hidden Markov Model in speech recognition

Jazyk

en

Originální abstrakt

In Hidden Markov models, speech data are modeled by Gaussian distributions. In this paper, we propose to Gaussianize the features to better fit to this modeling. A distribution of the data is estimated and a transform function is derived. We test three methods of the transform estimation (global, speaker based, frame based) and report results on the SPINE 2000 task with Sphinx recognizer. We conclude that the proposed method is a cheap way to increase the recognition accuracy.

Anglický abstrakt

In Hidden Markov models, speech data are modeled by Gaussian distributions. In this paper, we propose to Gaussianize the features to better fit to this modeling. A distribution of the data is estimated and a transform function is derived. We test three methods of the transform estimation (global, speaker based, frame based) and report results on the SPINE 2000 task with Sphinx recognizer. We conclude that the proposed method is a cheap way to increase the recognition accuracy.

BibTex


@inproceedings{BUT5410,
  author="Pavel {Matějka} and Milan {Sigmund} and Jan {Černocký}",
  title="Data adaptation for Hidden Markov Model in speech recognition",
  annote="In Hidden Markov models, speech data are modeled by Gaussian distributions. In this paper, we propose to Gaussianize the features to better fit to this modeling. A distribution of the data is estimated and a transform function is derived. We test three methods of the transform estimation (global, speaker based, frame based) and report results on the SPINE 2000 task with Sphinx recognizer. We conclude that the proposed method is a cheap way to increase the recognition accuracy.",
  booktitle="Proceedings of 8th Conference STUDENT EEICT 2002",
  chapter="5410",
  year="2002",
  month="april",
  pages="317",
  type="conference paper"
}