Advanced search
1 file | 604.96 KB

Large vocabulary continuous speech recognition with reservoir-based acoustic models

Fabian Triefenbach (UGent) , Kris Demuynck (UGent) and Jean-Pierre Martens (UGent)
(2014) IEEE SIGNAL PROCESSING LETTERS. 21(3). p.311-315
Author
Organization
Abstract
Thanks to recent research in neural network based acoustic modeling, Large Vocabulary Continuous Speech Recognition (LVCSR) has very recently made a significant step forwards. In search for further progress, the present paper investigates Reservoir Computing (RC) as an alternative new paradigm for acoustic modeling. This paradigm unifies the appealing dynamical modeling capacity of a recurrent neural network (RNN) with the simplicity and robustness of linear regression as a model for training the weights of that network. In previous work, an RC-HMM hybrid yielding good phone recognition accuracy on TIMIT could be designed, but no proof was offered that this success would also transfer to LVCSR. This paper describes the development of an RC-HMM hybrid that offers good recognition on the Wall Street Journal benchmark. For the 5K word task, word error rates on the Nov92 evaluation set are as low as 6.2% (bigram language model) and 3.9% (trigram). Given that RC-based acoustic modeling is a fairly new approach, these results open up very promising perspectives.
Keywords
Acoustic Modeling, Recurrent Neural Networks, Reservoir Computing, Large Vocabulary Continuous Speech Recognition, SHORT-TERM-MEMORY, NETS

Downloads

  • (...).pdf
    • full text
    • |
    • UGent only
    • |
    • PDF
    • |
    • 604.96 KB

Citation

Please use this url to cite or link to this publication:

Chicago
Triefenbach, Fabian, Kris Demuynck, and Jean-Pierre Martens. 2014. “Large Vocabulary Continuous Speech Recognition with Reservoir-based Acoustic Models.” Ieee Signal Processing Letters 21 (3): 311–315.
APA
Triefenbach, F., Demuynck, K., & Martens, J.-P. (2014). Large vocabulary continuous speech recognition with reservoir-based acoustic models. IEEE SIGNAL PROCESSING LETTERS, 21(3), 311–315.
Vancouver
1.
Triefenbach F, Demuynck K, Martens J-P. Large vocabulary continuous speech recognition with reservoir-based acoustic models. IEEE SIGNAL PROCESSING LETTERS. 2014;21(3):311–5.
MLA
Triefenbach, Fabian, Kris Demuynck, and Jean-Pierre Martens. “Large Vocabulary Continuous Speech Recognition with Reservoir-based Acoustic Models.” IEEE SIGNAL PROCESSING LETTERS 21.3 (2014): 311–315. Print.
@article{4234359,
  abstract     = {Thanks to recent research in neural network based acoustic modeling, Large Vocabulary Continuous Speech Recognition (LVCSR) has very recently made a significant step forwards. In search for further progress, the present paper investigates Reservoir Computing (RC) as an alternative new paradigm for acoustic modeling. This paradigm unifies the appealing dynamical modeling capacity of a recurrent neural network (RNN) with the simplicity and robustness of linear regression as a model for training the weights of that network. In previous work, an RC-HMM hybrid yielding good phone recognition accuracy on TIMIT could be designed, but no proof was offered that this success would also transfer to LVCSR. This paper describes the development of an RC-HMM hybrid that offers good recognition on the Wall Street Journal benchmark. For the 5K word task, word error rates on the Nov92 evaluation set are as low as 6.2\% (bigram language model) and 3.9\% (trigram). Given that RC-based acoustic modeling is a fairly new approach, these results open up very promising perspectives.},
  author       = {Triefenbach, Fabian and Demuynck, Kris and Martens, Jean-Pierre},
  issn         = {1070-9908},
  journal      = {IEEE SIGNAL PROCESSING LETTERS},
  keyword      = {Acoustic Modeling,Recurrent Neural Networks,Reservoir Computing,Large Vocabulary Continuous Speech Recognition,SHORT-TERM-MEMORY,NETS},
  language     = {eng},
  number       = {3},
  pages        = {311--315},
  title        = {Large vocabulary continuous speech recognition with reservoir-based acoustic models},
  url          = {http://dx.doi.org/10.1109/LSP.2014.2302080},
  volume       = {21},
  year         = {2014},
}

Altmetric
View in Altmetric
Web of Science
Times cited: