Publication Details
Fast Approximate Spoken Term Detection from Sequence of Phonemes
Szőke Igor, Ing., Ph.D. (DCGM FIT BUT)
Prasanna S.R.M. (IITG)
Heřmanský Hynek, prof. Ing., Dr.Eng. (DCGM FIT BUT)
Spoken term detection, probabilistic pronunciation model, phoneme recognition, confusion matrix
We investigate the detection of spoken terms in conversa-
tional speech using phoneme recognition with the objective
of achieving smaller index size as well as faster search speed.
Speech is processed and indexed as a sequence of one best
phoneme sequence. We propose the use of a probabilistic
pronunciation model for the search term to compensate for
the errors in the recognition of phonemes. This model is de-
rived using the pronunciation of the word and the phoneme
confusion matrix. Experiments are performed on the con-
versational telephone speech database distributed by NIST
for the 2006 spoken term detection. We achieve about 1500
times smaller index size and 14 times faster search speed
compared to the system using phoneme lattices, at the cost
of relatively lower detection performance.
@INPROCEEDINGS{FITPUB8670,
author = "Joel Pinto and Igor Sz\H{o}ke and S.R.M. Prasanna and Hynek He\v{r}mansk\'{y}",
title = "Fast Approximate Spoken Term Detection from Sequence of Phonemes",
pages = "28--33",
booktitle = "The 31st Annual International ACM SIGIR Conference 20-24 July 2008, Singapore",
year = 2008,
location = "Singapore, SG",
publisher = "Association for Computing Machinery",
ISBN = "978-90-365-2697-5",
language = "english",
url = "https://www.fit.vut.cz/research/publication/8670"
}