Conference paperHANNEMANN Mirko, TRMAL Jan, ONDEL Lucas, KESIRAJU Santosh and BURGET Lukáš. Bayesian jointsequence models for graphemetophoneme conversion. In: Proceedings of ICASSP 2017. New Orleans: IEEE Signal Processing Society, 2017, pp. 28362840. ISBN 9781509041176.  Publication language:  english 

Original title:  Bayesian jointsequence models for graphemetophoneme conversion 

Title (cs):  Bayesovské modelování sdružených sekvencí pro převod grafémů na fonémy 

Pages:  28362840 

Proceedings:  Proceedings of ICASSP 2017 

Conference:  42nd IEEE International Conference on Acoustics, Speech and Signal Processing 

Place:  New Orleans, US 

Year:  2017 

ISBN:  9781509041176 

DOI:  10.1109/ICASSP.2017.7952674 

Publisher:  IEEE Signal Processing Society 

URL:  http://www.fit.vutbr.cz/research/groups/speech/publi/2017/hannemann_icassp2017_0002836.pdf [PDF] 

Keywords 

Bayesian approach, jointsequence models,
weighted finite state transducers, lettertosound, graphemetophoneme conversion, hierarchical PitmanYorProcess 
Annotation 

This article is about Bayesian jointsequence models for graphemetophoneme conversion based on the jointsequence model (JSM).

Abstract 

We describe a fully Bayesian approach to graphemetophoneme
conversion based on the jointsequence model (JSM). Usually, standard
smoothed ngram language models (LM, e.g. KneserNey)
are used with JSMs to model graphone sequences (joint graphemephoneme
pairs). However, we take a Bayesian approach using a
hierarchical PitmanYorProcess LM. This provides an elegant alternative
to using smoothing techniques to avoid overtraining. No
heldout sets and complex parameter tuning is necessary, and several
convergence problems encountered in the discounted Expectation
Maximization (as used in the smoothed JSMs) are avoided. Every
step is modeled by weighted finite state transducers and implemented
with standard operations from the OpenFST toolkit. We
evaluate our model on a standard data set (CMUdict), where it gives
comparable results to the previously reported smoothed JSMs in
terms of phonemeerror rate while requiring a much smaller training/
testing time. Most importantly, our model can be used in a
Bayesian framework and for (partly) unsupervised training. 
BibTeX: 

@INPROCEEDINGS{
author = {Mirko Hannemann and Jan Trmal and Lucas Ondel and
Santosh Kesiraju and Luk{\'{a}}{\v{s}} Burget},
title = {Bayesian jointsequence models for
graphemetophoneme conversion},
pages = {28362840},
booktitle = {Proceedings of ICASSP 2017},
year = {2017},
location = {New Orleans, US},
publisher = {IEEE Signal Processing Society},
ISBN = {9781509041176},
doi = {10.1109/ICASSP.2017.7952674},
language = {english},
url = {http://www.fit.vutbr.cz/research/view_pub.php?id=11469}
} 
