Rule-Based Pronunciation Models to Handle OOV Words for Indonesian Automatic Speech Recognition System

A representative pronunciation dictionary becomes a necessity to cover large vocabulary in many domains. While creating a hand designed pronunciation dictionary in an extensive corpus is expensive and time-consuming, a rule-based pronunciation generation method is considerably effective to be applied in Indonesian language due to simpler graphemetophoneme mapping. In this paper, we propose a rule based G2P translation including an efficient way to handle abbreviation and word splitting. The accuracy of the proposed G2P translation achieved 93.65% for Indonesian word entries, and 87.29% for the overall entries that contain abbreviation and words from other language. The addition of rule based G2P translated OOV words to the dictionary has comparable speech recognition system performance with the hand designed dictionary. With the same amount of OOV addition, the speech recognition system performance is reduced by only 0.27 percent when using the rule-based G2P translated dictionary compared to the hand designed dictionary. From the experiment results, generating a dictionary for the OOV words with the proposed G2P mapping is considerably efficient without a big gap in the speech recognition performance score while reducing great amount of time and resources.