| Summary: | This thesis addresses the problem of modeling pronunciation variations in non-native English speech. In particular, it develops a computer assisted pronunciation learning (CAPL) system to assist speakers in Singapore to speak standard English. The dictionary of the CAPL system, also known as the lexicon, contains the sequences of sub-word units (usually phonemes) to describe how words are pronounced. However, it is often difficult to cover all the possible pronunciations. This work presents a method to improve a given initial lexicon to include new pronunciations that can explain the pronunciation variants of regional English accents in Singapore. The method learns pronunciation rules from an orthographically transcribed speech corpus to generate common pronunciation variants. All variants are then compiled into a compact pronunciation dictionary. The upgraded dictionary are then integrated into the CAPL system, where they are used to score the user's pronunciations. The work has three novel contributions. Firstly it constructs a Singapore English corpus, which is one of the few standard corpora for speech research on the regional accent. The corpus consists of sentences used in the standard LDC TIMIT corpus. Secondly, it learns pronunciation rules from the speech data using a combination of data-driven and knowledge-based approaches in pronunciation modeling. Thirdly, it designs a prototype pronunciation scoring algorithm to evaluate and score the goodness of pronunciation in the CAPL system. The simulation shows satisfactory performance in the proposed pronunciation scoring system.
|
|---|