Bayesian Learning of a Language Model from Continuous Speech

Graham NEUBIG; Masato MIMURA; Shinsuke MORI; Tatsuya KAWAHARA

doi:10.1587/transinf.E95.D.614

Bayesian Learning of a Language Model from Continuous Speech

Graham NEUBIG, Masato MIMURA, Shinsuke MORI, Tatsuya KAWAHARA

Full Text Views

0

Cite this

Summary :

We propose a novel scheme to learn a language model (LM) for automatic speech recognition (ASR) directly from continuous speech. In the proposed method, we first generate phoneme lattices using an acoustic model with no linguistic constraints, then perform training over these phoneme lattices, simultaneously learning both lexical units and an LM. As a statistical framework for this learning problem, we use non-parametric Bayesian statistics, which make it possible to balance the learned model's complexity (such as the size of the learned vocabulary) and expressive power, and provide a principled learning algorithm through the use of Gibbs sampling. Implementation is performed using weighted finite state transducers (WFSTs), which allow for the simple handling of lattice input. Experimental results on natural, adult-directed speech demonstrate that LMs built using only continuous speech are able to significantly reduce ASR phoneme error rates. The proposed technique of joint Bayesian learning of lexical units and an LM over lattices is shown to significantly contribute to this improvement.

Publication: IEICE TRANSACTIONS on Information Vol.E95-D No.2 pp.614-625

Publication Date: 2012/02/01

Publicized

Online ISSN: 1745-1361

DOI: 10.1587/transinf.E95.D.614

Type of Manuscript: PAPER

Category: Speech and Hearing

Cite this

Copy

Graham NEUBIG, Masato MIMURA, Shinsuke MORI, Tatsuya KAWAHARA, "Bayesian Learning of a Language Model from Continuous Speech" in IEICE TRANSACTIONS on Information, vol. E95-D, no. 2, pp. 614-625, February 2012, doi: 10.1587/transinf.E95.D.614.
Abstract: We propose a novel scheme to learn a language model (LM) for automatic speech recognition (ASR) directly from continuous speech. In the proposed method, we first generate phoneme lattices using an acoustic model with no linguistic constraints, then perform training over these phoneme lattices, simultaneously learning both lexical units and an LM. As a statistical framework for this learning problem, we use non-parametric Bayesian statistics, which make it possible to balance the learned model's complexity (such as the size of the learned vocabulary) and expressive power, and provide a principled learning algorithm through the use of Gibbs sampling. Implementation is performed using weighted finite state transducers (WFSTs), which allow for the simple handling of lattice input. Experimental results on natural, adult-directed speech demonstrate that LMs built using only continuous speech are able to significantly reduce ASR phoneme error rates. The proposed technique of joint Bayesian learning of lexical units and an LM over lattices is shown to significantly contribute to this improvement.
URL: https://global.ieice.org/en_transactions/information/10.1587/transinf.E95.D.614/_p

Copy

@ARTICLE{e95-d_2_614,
author={Graham NEUBIG, Masato MIMURA, Shinsuke MORI, Tatsuya KAWAHARA, },
journal={IEICE TRANSACTIONS on Information},
title={Bayesian Learning of a Language Model from Continuous Speech},
year={2012},
volume={E95-D},
number={2},
pages={614-625},
abstract={We propose a novel scheme to learn a language model (LM) for automatic speech recognition (ASR) directly from continuous speech. In the proposed method, we first generate phoneme lattices using an acoustic model with no linguistic constraints, then perform training over these phoneme lattices, simultaneously learning both lexical units and an LM. As a statistical framework for this learning problem, we use non-parametric Bayesian statistics, which make it possible to balance the learned model's complexity (such as the size of the learned vocabulary) and expressive power, and provide a principled learning algorithm through the use of Gibbs sampling. Implementation is performed using weighted finite state transducers (WFSTs), which allow for the simple handling of lattice input. Experimental results on natural, adult-directed speech demonstrate that LMs built using only continuous speech are able to significantly reduce ASR phoneme error rates. The proposed technique of joint Bayesian learning of lexical units and an LM over lattices is shown to significantly contribute to this improvement.},
keywords={},
doi={10.1587/transinf.E95.D.614},
ISSN={1745-1361},
month={February},}

Copy

TY - JOUR
TI - Bayesian Learning of a Language Model from Continuous Speech
T2 - IEICE TRANSACTIONS on Information
SP - 614
EP - 625
AU - Graham NEUBIG
AU - Masato MIMURA
AU - Shinsuke MORI
AU - Tatsuya KAWAHARA
PY - 2012
DO - 10.1587/transinf.E95.D.614
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E95-D
IS - 2
JA - IEICE TRANSACTIONS on Information
Y1 - February 2012
AB - We propose a novel scheme to learn a language model (LM) for automatic speech recognition (ASR) directly from continuous speech. In the proposed method, we first generate phoneme lattices using an acoustic model with no linguistic constraints, then perform training over these phoneme lattices, simultaneously learning both lexical units and an LM. As a statistical framework for this learning problem, we use non-parametric Bayesian statistics, which make it possible to balance the learned model's complexity (such as the size of the learned vocabulary) and expressive power, and provide a principled learning algorithm through the use of Gibbs sampling. Implementation is performed using weighted finite state transducers (WFSTs), which allow for the simple handling of lattice input. Experimental results on natural, adult-directed speech demonstrate that LMs built using only continuous speech are able to significantly reduce ASR phoneme error rates. The proposed technique of joint Bayesian learning of lexical units and an LM over lattices is shown to significantly contribute to this improvement.
ER -