Grapheme-to-phoneme conversion methods for minority language conditions

This study attempts to investigate the grapheme-to-phoneme conversion approaches for minority language conditions. Instead of isolated-word data for major languages, sentence-form data is defined to be a proper form of training data for minority languages. Joint-multigram Model and Hidden Markov Model were examined in this study. The “treat-sentence-as-word” training method and the forced-alignment process were proposed to extend the Joint-multigram Model and the Hidden Markov Model respectively to meet the minority language conditions. Results get from the sentence-form training data using our proposed methods are as good as the results get from the isolated-word training data using previous proposed methods. The Joint-multigram Model performs better for well-designed training data, while the Hidden Markov Model has more error capacity and is more proper for minority language conditions.