Conference paper

GRÉZL František, EGOROVA Ekaterina and KARAFIÁT Martin. Study of Large Data Resources for Multilingual Training and System Porting. In: Procedia Computer Science. Yogyakarta: Elsevier Science, 2016, pp. 15-22. ISSN 1877-0509. Available from: http://www.sciencedirect.com/science/article/pii/S1877050916300382
Publication language:english
Original title:Study of Large Data Resources for Multilingual Training and System Porting
Title (cs):Studie velkých datových zdrojů pro multilingvální trénování a portování systémů
Pages:15-22
Proceedings:Procedia Computer Science
Conference:The 5th International Workshop on Spoken Language Technologies for Under-resourced Languages (SLTU'16)
Place:Yogyakarta, ID
Year:2016
URL:http://www.sciencedirect.com/science/article/pii/S1877050916300382
Journal:Procedia Computer Science, Vol. 2016, No. 81, CZ
ISSN:1877-0509
Publisher:Elsevier Science
URL:http://www.fit.vutbr.cz/research/groups/speech/publi/2016/grezl_sltu2016_08-8028.pdf [PDF]
Files: 
+Type Name Title Size Last modified
icongrezl_sltu2016_08-8028.pdf246 KB2016-07-12 16:16:21
^ Select all
With selected:
Keywords
Stacked Bottle-Neck; feature extraction; multilingual training; large data; Fisher database
Annotation
This study investigates the behavior of a feature extraction neural network model trained on a large amount of single language data ("source language") on a set of under-resourced target languages. The coverage of the source language acoustic space was changed in two ways: (1) by changing the amount of training data and (2) by altering the level of detail of acoustic units (by changing the triphone clustering). We observe the effect of these changes on the performance on target language in two scenarios: (1) the source-language NNs were used directly, (2) NNs were first ported to target language. The results show that increasing coverage as well as level of detail on the source language improves the target language system performance in both scenarios. For the first one, both source language characteristic have about the same effect. For the second scenario, the amount of data in source language is more important than the level of detail. The possibility to include large data into multilingual training set was also investigated. Our experiments point out possible risk of over-weighting the NNs towards the source language with large data. This degrades the performance on part of the target languages, compared to the setting where the amounts of data per language are balanced.
Abstract
This study investigates the behavior of a feature extraction neural network model trained on a large amount of single language data ("source language") on a set of under-resourced target languages. The coverage of the source language acoustic space was changed in two ways: (1) by changing the amount of training data and (2) by altering the level of detail of acoustic units (by changing the triphone clustering). We observe the effect of these changes on the performance on target language in two scenarios: (1) the source-language NNs were used directly, (2) NNs were first ported to target language. The results show that increasing coverage as well as level of detail on the source language improves the target language system performance in both scenarios. For the first one, both source language characteristic have about the same effect. For the second scenario, the amount of data in source language is more important than the level of detail. The possibility to include large data into multilingual training set was also investigated. Our experiments point out possible risk of over-weighting the NNs towards the source language with large data. This degrades the performance on part of the target languages, compared to the setting where the amounts of data per language are balanced.
BibTeX:
@INPROCEEDINGS{
   author = {Franti{\v{s}}ek Gr{\'{e}}zl and Ekaterina Egorova and Martin
	Karafi{\'{a}}t},
   title = {Study of Large Data Resources for Multilingual Training and
	System Porting},
   pages = {15--22},
   booktitle = {Procedia Computer Science},
   journal = {Procedia Computer Science},
   volume = {2016},
   number = {81},
   year = {2016},
   location = {Yogyakarta, ID},
   publisher = {Elsevier Science},
   ISSN = {1877-0509},
   language = {english},
   url = {http://www.fit.vutbr.cz/research/view_pub.php.en?id=11222}
}

Your IPv4 address: 54.81.6.121
Switch to IPv6 connection

DNSSEC [dnssec]