Training LVCSR systems on thousands of hours of data (2005)
| Venue: | IN: PROC. ICASSP |
| Citations: | 18 - 5 self |
BibTeX
@INPROCEEDINGS{Evermann05traininglvcsr,
author = {G. Evermann and H. Y. Chan and M. J. F. Gales and B. Jia and D. Mrva and P. C. Woodland and K. Yu},
title = {Training LVCSR systems on thousands of hours of data},
booktitle = {IN: PROC. ICASSP},
year = {2005},
pages = {209--212},
publisher = {}
}
Bookmark
 |
 |
 |
 |
 |
 |
OpenURL
Abstract
Typical systems for large vocabulary conversational speech recognition (LVCSR) have been trained on a few hundred hours of carefully transcribed acoustic training data. This paper describes an LVCSR system for the conversational telephone speech (CTS) task trained on more than 2000 hours of data for which only approximate transcriptions were available. The challenges of dealing which such a large data set and the accuracy improvements over the small baseline system are discussed. The effect on both acoustic and language modelling performance is studied. Overall increasing the training data size from 360h to 2200h and optimising the training procedure reduced the word error rate on the DARPA/NIST 2003 eval set by about 20 % relative.
Citations
| 88 | Entropy-based pruning of backoff language models - Stolcke - 1998 |
| 36 | Getting More Mileage from Web Text Sources for Conversational Speech Language Modeling using Class-Dependent Mixtures - Bulyko, Ostendorf, et al. - 2003 |
| 11 | Development of the 2003 CU-HTK conversational telephone speech system - Evermann, Chan, et al. |
| 11 | Investigation of acoustic modelling techniques for LVCSR systems - Liu, Gales, et al. - 2005 |
| 8 | Design of Fast LVCSR Systems - Evermann, Woodland - 2003 |
| 7 | From Switchboard to Fisher: Telephone Collection Protocols, their Uses and Yields - Cieri, Miller, et al. - 2003 |
| 4 | Shared resources for robust speech-to-text technology - Strassel, Miller, et al. - 2003 |
| 1 | From CTRAN to WordWave: More CTS for less - Kimball, Iyer, et al. |
