Comparing manually-developed and data-driven rules for P2P learning

Abstract

Phoneme-to-phoneme (P2P) learning provides a mechanism for predicting the pronunciation of a word based on its pronunciation in a different accent, dialect or language. The authors evaluate the effectiveness of manually-developed as well as automatically derived P2P rules for British to South African English pronunciation conversion. Using the freely-available Oxford Advanced Learners Dictionary of Contemporary English (OALD) as source, the two approaches to P2P conversion are compared to a manually-developed South African English pronunciation dictionary. The authors show that, when the British English pronunciation is known, a small manually-derived rule set is able to approximate the South African pronunciation surprisingly well. Furthermore they demonstrate that the best performance is achieved by data-driven P2P learning, which proves to be a better mechanism for pronunciation prediction than both manually-derived P2P rules as well as data-driven grapheme-to-phoneme (G2P) conversion.