INDONESIAN-ENGLISH CODE-SWITCHING HANDLING IN INDONESIAN AUTOMATIC SPEECH RECOGNITION SYSTEM

• Main Author: Hartanto, Roland
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39993
• Institution: Institut Teknologi Bandung

English as an international language greatly influences the use of Indonesian. Indonesian-English code-switching is one of the effects that can be felt. Codeswitching is the event of using a language along with at least one other language. This causes performance degradation of the automatic speech recognition system because of unrecognized words and phonemes. An automatic speech recognition system for the Indonesian-English code-switching was built. The handling of acoustic modeling is carried out using the English to Indonesian phoneme mapping approach, rule-based phone merging using the universal phonemes (International Phonetic Alphabet), and data-driven approach by measuring the Bhattacharyya distance of phonemes. In the data-driven approach, clustering of English and Indonesian phonemes is conducted using the agglomerative clustering approach. The making of an acoustic model uses the Hidden Markov Model (HMM) by calculating the emission probability using the Gaussian Mixture Model (GMM). The handling of language modeling is done by increasing the number of text corpus by augmenting Indonesian sentences containing English words chosen based on the frequency of occurrence of English words. The evaluation results showed a decrease in word error rate (WER) of 1.37% from the handling of the acoustic model using the phone merging method with an agglomerative clustering and 2.05% for handling language models relative to the baseline system. The decrease in WER was also accompanied by a significant reduction in English word recognition errors without reducing the performance of Indonesian language recognition.