Automatic speech recognition tasks whose domains involve recognizing proper names are usually challenging due to the infinite number of possible lexical items. In some tasks such as automated directories, to avoid unconstrained recognitions, one could design the system to handle a finite number of proper names at a time. However, the recognition grammars need to be generated in unsupervised fashions. In this paper, a recognition system capable of selecting a Thai proper name from dynamic name lists was proposed and evaluated upon name lists of various sizes. The system decomposes names to sequences of Grams, lexical units defined with specific pronunciations and spellings, upon which pronunciation grammars are produced and utilized as recognition constraints. Recognition accuracies of above 85% to approximately 91% were obtained on lists containing 10 to 80 names.
[1]
Atiwong Suchato,et al.
Romanization of Thai Proper Names Based on Popularity of Usages
,
2009,
PAKDD.
[2]
Wanchai Rivepiboon,et al.
A Unified Model of Thai Romanization and Word Segmentation
,
2004,
PACLIC.
[3]
Wirote Aroonmanakun,et al.
Collocation and Thai Word Segmentation
,
2002
.
[4]
Biing-Hwang Juang,et al.
Hidden Markov Models for Speech Recognition
,
1991
.
[5]
A. Suchato,et al.
Phoneme Classification Study for Thai Segment-Based Acoustic Models
,
2006,
2006 International Symposium on Communications and Information Technologies.