Spelling Recognition for Two Commonly-Used Thai Spelling Methods

C. Pisarn, T. Theeramunkong (Thailand), and N. Cercone (Canada)

Keywords

Spelling recognition, HMMs, Continuous speech recognition, Speaker independent, Opentest environment.

Abstract

Since it is rare to achieve 100% speech recognition, spelling recognition is a practical and efficient way to assist an automatic speech recognition system (ASR). Unlike other languages, Thai has several word spelling styles. Two common spelling styles are introduced and discussed. Spelling recognition systems of each spelling style are established using two small spelling corpora and the NECTEC-ATR Thai continuous speech corpus. An open-test and speaker independent system is designed to offer impartial comparisons. The recognition results show that for the easier spelling style, the recognizer trained by the NECTEC-ATR can gain accuracy up to 85.26% on the mixed-type environment with a bigram language model. However, for both two spelling styles, the recognizers trained by a small spelling-speech corpus outperforms the recognizer trained by the larger general speech corpus, i.e. the NECTEC-ATR Thai speech corpus. The best accuracies for the two spelling styles are 89.08% and 70.08% when a mixed-type environment bigram language model is applied.

Important Links:



Go Back