Automatic Extraction of Pronunciation Rules and Generating Pronunciation Variants of Words to Use in Automatic Continuous Speech Recognition

Automatic Extraction of Pronunciation Rules and Generating Pronunciation Variants of Words to Use in Automatic Continuous Speech Recognition

Abstract

Words are pronounced in various ways in continuous speech. So the lexicon of a continuous speech recognition system is better to contain various pronunciations of each word. The accuracy of word recognition will be improved in this way. An automatic method to generate pronunciation variants of words is introduced in this paper. Pronunciation rules are learned by comparing aligned pairs of reference and recognized phonetic transcriptions of words in this method. In addition, some knowledge-based rewrite rules are added to previous list of learned rules. Consequently after statistical pruning of rules, we have used them to generate pronunciation variants of words by applying them to phonemic transcriptions of words. This method has many advantages in comparison to adding pronunciation variants manually, as it takes into account errors of phone recognizer system and computes application likelihood of each pronunciation variant automatically. This method is implemented by using FARSDAT Persian speech corpus. By usage of generated pronunciation variants in the lexicon of SHENAVA a Persian ACSR, an improvement of as high as 3.47% is achieved in words recognition accuracy.

References